WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India Towards a Theory Model for Product Search Beibei li Anindya Ghose Panagiotis G. Ipeirotis bli@stern. nyu. edu lose@stern. nyu. edu panos @stern. nyu. edu Department of Information, Operations, and Management Sciences Leonard N. Stern School of Business, New York University New York. New York 10012. USA ABSTRACT a product is different from the process of finding With the growing pervasiveness of the Internet, online search document or object. Customers do not simply for products and services is constantI easing. Most product something relevant to their search, but also try to search engines are based on adaptations of theoretical models " best"deal that satisfies their specific desired criteria it is difficult 四 devised for information retrieval. However. the decision mecha- ntify the notion of"best" prod nism that underlies the process of buying a product is different trying to understand what the users are optimizing than the process of locating relevant documents or objects Today's product search engines provide only rudimentary We propose a theory model for product search based on ranking facilities for search results, typically using a single rank expected utility theory from economics. Specifically, we propose ing criterion such as name, price, best selling(volume of sales a ranking technique in which we rank highest the products that or more recently, using customer review ratings. This approach enerate the highest surplus, after the purchase. In a sense. the has quite a few shortcor multidimen- top ranked products are the"best value for money for a specif sional preferences of consumers. Second, it fails to leverage Pat. Our approach builds on research on"demand estimation" the information generated by the online communities, goin beyond simple numerical ratings. Third, it hardly account which further research can build on. We build algorithms that for the heterogeneity of consumers. These drawbacks highl ake into account consumer demographics, heterogeneity of necessitate a recommendation strategy for products that can consumer preferences, and also account for the varying price of better model consumers' underlying behavior, to capture their the products. We show how to achieve this without knowing the multidimensional prefe es and heterogeneous tastes. demographics or purchasing histories of individual consumers Recommender systems [1] could fix some of these problems but by using aggregate demand data. We evaluate our work. but, to the best of our knowledge, existing techniques still have by applying the techniques on hotel search. Our extensive limitations. First, most recommendation mechanisms require ser studies, using more than 15,000 user-provided ranking consumers' to log into the system. However, in reality many comparisons, demonstrate an overwhelming preference for the consumers browse only anonymously. Due to the lack of any rankings generated by our techniques, compared to a large eaningful, personalized recommendat consumers do no number of existing strong state-of-the-art baselines feel compelled to login before purchasing. For example, or Travelocity, less than 2% of the users login. But even when the Categories and Subject Descriptors login, before or after a purchase, consumers are reluctant to H.3.3(Information Storage and Retrieval]: Information of reasons (e. g, time constraints or privacy issues). Therefore, most context information is missing at the individual consumer General terms level. Second, for goods with a low purchase frequency for an Algorithms, Economics, Experimentation, Measurement individual consumer such as hotels. cars. real estate. or even electronics, there are few repeated purchases we could leverage Keywords towards building a predictive model (i.e, models based on onsumer Surplus, Economics, Product Search, Ranking. Text collaborative filtering). Third, and potentially more importantly, Mining, User-Generated Content, Utility Theory as privacy issues become increasingly important, marketers may not be able to observe the individual-level purchase history of each consumer(or consumer segment). In contrast, aggregate 1. INTRODUCTION purchase statistics(e. g, market share) sier to obtain. As a Online search for products is increasing in popularity, as more consequence, many algorithms that rely on knowing individual and more users search and purchase products from the Internet level behavior lack the ability of deriving consumer preferences lost search engines for products today are based on models of from such aggregate data. relevance from"classic"information retrieval theory [22] or use Alternative techniques try to identify the"Pareto optimal variants of faceted search[27] to facilitate browsing. However, set of results [3]. Unfortunately, the feasibility of this approach the decision mechanism that underlies the process of buying diminishes as the number of product characteristics With more than five haracteristics, the probability of a This work Ipported by NSF grants IIs-0613847 and IIS-06413846 point being classified as"Pareto optimal "dramatically increases. is held by the International World wide Web Conference Com- As a consequence, the set of Pareto optimal results soon includes nittee(Iw3C2). Distribution of these papers is limited to classroom use, prod So, how to generate the "best"ranking of products wh 011, March 28-April 1, 2011, Hyderabad, India. ACM978-1-4503-0632-4/1103
Towards a Theory Model for Product Search∗ Beibei Li bli@stern.nyu.edu Anindya Ghose aghose@stern.nyu.edu Panagiotis G. Ipeirotis panos@stern.nyu.edu Department of Information, Operations, and Management Sciences Leonard N. Stern School of Business, New York University New York, New York 10012, USA ABSTRACT With the growing pervasiveness of the Internet, online search for products and services is constantly increasing. Most product search engines are based on adaptations of theoretical models devised for information retrieval. However, the decision mechanism that underlies the process of buying a product is different than the process of locating relevant documents or objects. We propose a theory model for product search based on expected utility theory from economics. Specifically, we propose a ranking technique in which we rank highest the products that generate the highest surplus, after the purchase. In a sense, the top ranked products are the “best value for money” for a specific user. Our approach builds on research on “demand estimation” from economics and presents a solid theoretical foundation on which further research can build on. We build algorithms that take into account consumer demographics, heterogeneity of consumer preferences, and also account for the varying price of the products. We show how to achieve this without knowing the demographics or purchasing histories of individual consumers but by using aggregate demand data. We evaluate our work, by applying the techniques on hotel search. Our extensive user studies, using more than 15,000 user-provided ranking comparisons, demonstrate an overwhelming preference for the rankings generated by our techniques, compared to a large number of existing strong state-of-the-art baselines. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval General Terms Algorithms, Economics, Experimentation, Measurement Keywords Consumer Surplus, Economics, Product Search, Ranking, Text Mining, User-Generated Content, Utility Theory 1. INTRODUCTION Online search for products is increasing in popularity, as more and more users search and purchase products from the Internet. Most search engines for products today are based on models of relevance from “classic” information retrieval theory [22] or use variants of faceted search [27] to facilitate browsing. However, the decision mechanism that underlies the process of buying ∗This work was supported by NSF grants IIS-0643847 and IIS-0643846. Copyright is held by the International World Wide Web Conference Committee (IW3C2). Distribution of these papers is limited to classroom use, and personal use by others. WWW 2011, March 28–April 1, 2011, Hyderabad, India. ACM 978-1-4503-0632-4/11/03. a product is different from the process of finding a relevant document or object. Customers do not simply seek to find something relevant to their search, but also try to identify the “best” deal that satisfies their specific desired criteria. Of course, it is difficult to quantify the notion of “best” product without trying to understand what the users are optimizing. Today’s product search engines provide only rudimentary ranking facilities for search results, typically using a single ranking criterion such as name, price, best selling (volume of sales), or more recently, using customer review ratings. This approach has quite a few shortcomings. First, it ignores the multidimensional preferences of consumers. Second, it fails to leverage the information generated by the online communities, going beyond simple numerical ratings. Third, it hardly accounts for the heterogeneity of consumers. These drawbacks highly necessitate a recommendation strategy for products that can better model consumers’ underlying behavior, to capture their multidimensional preferences and heterogeneous tastes. Recommender systems [1] could fix some of these problems but, to the best of our knowledge, existing techniques still have limitations. First, most recommendation mechanisms require consumers’ to log into the system. However, in reality many consumers browse only anonymously. Due to the lack of any meaningful, personalized recommendations, consumers do not feel compelled to login before purchasing. For example, on Travelocity, less than 2% of the users login. But even when they login, before or after a purchase, consumers are reluctant to give their individual demographic information due to a variety of reasons (e.g., time constraints or privacy issues). Therefore, most context information is missing at the individual consumer level. Second, for goods with a low purchase frequency for an individual consumer, such as hotels, cars, real estate, or even electronics, there are few repeated purchases we could leverage towards building a predictive model (i.e., models based on collaborative filtering). Third, and potentially more importantly, as privacy issues become increasingly important, marketers may not be able to observe the individual-level purchase history of each consumer (or consumer segment). In contrast, aggregate purchase statistics (e.g., market share) are easier to obtain. As a consequence, many algorithms that rely on knowing individuallevel behavior lack the ability of deriving consumer preferences from such aggregate data. Alternative techniques try to identify the “Pareto optimal” set of results [3]. Unfortunately, the feasibility of this approach diminishes as the number of product characteristics increases. With more than five or six characteristics, the probability of a point being classified as “Pareto optimal” dramatically increases. As a consequence, the set of Pareto optimal results soon includes every product. So, how to generate the “best” ranking of products when WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 327
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad India multiple criteria? For this purpose, we use two 2. THEORY MODEL fundamental concepts from economics: utility and surplus. Util- In this section. we ned as a measure of the relative satisfaction desirability of, consumption of various goods and services [17]. by formalizing our problem and introducing the "economic Each product provides consumers with an overall utility, which view"of consumer rational choices. For better understanding we introduce the following theoretical bases: utility theory. individual product characteristics. At the same time, the action of purchasing deprives the customer from the utility of the ney that is spent for buying the product. With the assump- 2.1 Problem Description tion consumers rationality, the de making process behind n general, our main goal is to identify the best products for a purchasing can be viewed as a process of utility maximization that takes into consideration both product quality and price Based on utility theory, we propose to design a new ranking ExAMPLE 1. Alice is looking for a hotel in New York City system that uses demand-estimation approaches from economics She prefers a place of good quality but preferably costing not o generate the weights that consumers implicitly assign to each more than $300 per night. She conducts a faceted search(e.g. individual product characteristic. An important characteristic of with respect to price and ratings ): Unfortunately, with explicit this approach is that it does not require purchasing information price constrain, she may miss some "great deal"with much individual customers but rather relies on aggregate demand higher value but a slightly higher price. For instance, the 5-star data. Based on the estimated weights, we then derive the surplus Mandarin hotel happens to run a promotion that week with a for each product, which represents how much extra utility one discounted price of $333 per night. With the most luxurious can obtain by purchasing a product. Finally, we rank all the environment and room services, the price for Mandarin would products ording to their surplus. We extend our ranking normally be around S900 per night otherwise. So, although the strategy to a personalized level, based on the distribution of price is $33 aboue her budget, Alice would certainly be willing to "grab the deal"if this hotel appeared in the search result We instantiate our research by building a search engine for otels, based on a unique data set containing transactions from However, the problem is how can Alice know that such a Nov. 2008 to Jan. 2009 for US hotels from a major travel deal exists? In other words, how can we improve the search so web site. Our extensive user studies, using more than 15000 that it can help Alice identify the "best value" products? To evaluations, demonstrate an overwhelming preference for examine this problem, we introduce the concept of surplus from economics. It is a number of existing strong baselines. from the exchange of goods [17]. If we can derive the surplus he following from each product, then by ranking the products according to We aim at making recommendations based on better un their surplus, we can easily find the best product that provides derstanding of the underlying"causality"of consumers' the highest benefit purchase decisions. We present a user model that captures derive the surplus so that we can quantify the gain from buying the decision-making process of consumers, leading to a a product? To do so, we introduce another concept: utility. contrast to building a"black-box"style predictive model 2.2 Choice Decisions and Utility Maximization using machine learning algorithms. The causal model re- Surplus can be derived from utility and rational choice the- laxes the assumption of a"consistent environment"across ories. A fundamental notion in utility theory is that each training and testing data sets and allows for changes in the consumer is endowed with an associated utility function U, odeling environment and predicts what should happen which is "a measure of the satisfaction from consumption of n when things chan various goods and services"[17. The rationality assumption We infer personal preferences from aggregate data, in a defines that each person tries to maximize its own utility privacy-preserving manner. Our algorithm learns con- In the context of purchasing decisions, we assume tI sumer preferences based on the largely anonymous, pub- consumer has access to a set of products, each product licly observed distributions of consumer demographics as a price. Informally, buying a product involves the exo well as the observed aggregate-level purchases (i.e, anony- of money for a product. Therefore, to analyze the purchasing mous purchases and market shares in NYC and LA), not behavior we need two components for the utility function ning from the identified behavior or demographics Utility of Product: The utility that the consumer will gain of each individual We propose a ranking method using the notion of sur- by buying the product, and plus, which is not only theory-driven but also generates Utility of Money: The utility that the consumer will lose systematically better results than current approaches by paying the price for that product In general, a consumer buys the product that maximizes utility. We present an extensive experimental study: using six and does so only if the utility gained by purchasing the product hotel markets, 15000 user evaluations, and using blind ests, we demonstrate that the generated rankings is higher than the corresponding, lost, utility of the money. significantly better than existing approaches The rest of the paper is organized as follows. Section 2 n products, and each product X, has a price Pi 2 Before the the background. Section 3 explains how we estimate the model purchase, the consumer has some disposable income I that ly how we compute the weights associat with product characteristics. Section 4 discusses how we build modeling framework that we adop our basic rankings, and how we can personalize the presente the experimental evaluation, even imperfect theories generate good esults. Section 5 provides the setting for the experiment imental results valuation,and Section 6 discusses the results. Finally, Section 7 To allow for the possibility of not buying anything, we also add a dummy product Xo with price po = 0, which corresponds to the choice of discusses related work and Section 8 concludes buying anything. 328
consumers use multiple criteria? For this purpose, we use two fundamental concepts from economics: utility and surplus. Utility is defined as a measure of the relative satisfaction from, or desirability of, consumption of various goods and services [17]. Each product provides consumers with an overall utility, which can be represented as the aggregation of weighted utilities of individual product characteristics. At the same time, the action of purchasing deprives the customer from the utility of the money that is spent for buying the product. With the assumption consumers rationality, the decision-making process behind purchasing can be viewed as a process of utility maximization that takes into consideration both product quality and price. Based on utility theory, we propose to design a new ranking system that uses demand-estimation approaches from economics to generate the weights that consumers implicitly assign to each individual product characteristic. An important characteristic of this approach is that it does not require purchasing information for individual customers but rather relies on aggregate demand data. Based on the estimated weights, we then derive the surplus for each product, which represents how much extra utility one can obtain by purchasing a product. Finally, we rank all the products according to their surplus. We extend our ranking strategy to a personalized level, based on the distribution of consumers’ demographics. We instantiate our research by building a search engine for hotels, based on a unique data set containing transactions from Nov. 2008 to Jan. 2009 for US hotels from a major travel web site. Our extensive user studies, using more than 15000 user evaluations, demonstrate an overwhelming preference for the ranking generated by our techniques, compared to a large number of existing strong baselines. The major contributions of our research are the following: • We aim at making recommendations based on better understanding of the underlying “causality” of consumers’ purchase decisions. We present a user model that captures the decision-making process of consumers, leading to a better understanding of consumer preferences. This is in contrast to building a “black-box” style predictive model using machine learning algorithms. The causal model relaxes the assumption of a “consistent environment” across training and testing data sets and allows for changes in the modeling environment and predicts what should happen even when things change. • We infer personal preferences from aggregate data, in a privacy-preserving manner. Our algorithm learns consumer preferences based on the largely anonymous, publicly observed distributions of consumer demographics as well as the observed aggregate-level purchases (i.e., anonymous purchases and market shares in NYC and LA), not by learning from the identified behavior or demographics of each individual. • We propose a ranking method using the notion of surplus, which is not only theory-driven but also generates systematically better results than current approaches. • We present an extensive experimental study: using six hotel markets, 15000 user evaluations, and using blind tests, we demonstrate that the generated rankings are significantly better than existing approaches. The rest of the paper is organized as follows. Section 2 gives the background. Section 3 explains how we estimate the model parameters, specifically how we compute the weights associated with product characteristics. Section 4 discusses how we build our basic rankings, and how we can personalize the presented results. Section 5 provides the setting for the experimental evaluation, and Section 6 discusses the results. Finally, Section 7 discusses related work and Section 8 concludes. 2. THEORY MODEL In this section, we provide the background economic theory that explains the basic concepts behind our model. We start by formalizing our problem and introducing the “economic view” of consumer rational choices. For better understanding, we introduce the following theoretical bases: utility theory, characteristics-based theory, and surplus. 2.1 Problem Description In general, our main goal is to identify the best products for a consumer. The example illustrates this: Example 1. Alice is looking for a hotel in New York City. She prefers a place of good quality but preferably costing not more than $300 per night. She conducts a faceted search (e.g., with respect to price and ratings): Unfortunately, with explicit price constrain, she may miss some “great deal” with much higher value but a slightly higher price. For instance, the 5-star Mandarin hotel happens to run a promotion that week with a discounted price of $333 per night. With the most luxurious environment and room services, the price for Mandarin would normally be around $900 per night otherwise. So, although the price is $33 above her budget, Alice would certainly be willing to ”grab the deal” if this hotel appeared in the search result. However, the problem is how can Alice know that such a deal exists? In other words, how can we improve the search so that it can help Alice identify the “best value” products? To examine this problem, we introduce the concept of surplus from economics. It is a measure of the benefits consumers derive from the exchange of goods [17]. If we can derive the surplus from each product, then by ranking the products according to their surplus, we can easily find the best product that provides the highest benefits to a consumer. Now the question is, how to derive the surplus so that we can quantify the gain from buying a product? To do so, we introduce another concept: utility. 2.2 Choice Decisions and Utility Maximization Surplus can be derived from utility and rational choice theories. A fundamental notion in utility theory is that each consumer is endowed with an associated utility function U, which is “a measure of the satisfaction from consumption of various goods and services” [17]. The rationality assumption defines that each person tries to maximize its own utility.1 In the context of purchasing decisions, we assume that the consumer has access to a set of products, each product having a price. Informally, buying a product involves the exchange of money for a product. Therefore, to analyze the purchasing behavior we need two components for the utility function: • Utility of Product: The utility that the consumer will gain by buying the product, and • Utility of Money: The utility that the consumer will lose by paying the price for that product. In general, a consumer buys the product that maximizes utility, and does so only if the utility gained by purchasing the product is higher than the corresponding, lost, utility of the money. More formally, assume that the consumer has a choice across n products, and each product Xj has a price pj . 2 Before the purchase, the consumer has some disposable income I that 1While in reality consumers are not always rational, it is a convenient modeling framework that we adopt in this paper. As we demonstrate in the experimental evaluation, even imperfect theories generate good experimental results. 2To allow for the possibility of not buying anything, we also add a dummy product X0 with price p0 = 0, which corresponds to the choice of not buying anything. WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 328
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India generates a money utility Um(I). The decision to →U0-u0-p)-a X generates a product utility Up(X,) and, simt paying the price P; decreases the money utility to Un Assuming that the consumer strives to optimize its ov the purchased product X, is the one that gives the highest This approach naturally generates a ranking order for the products: The products that generate the highest increase in utility should be ranked on top. Thus, to compute the in utility, we need the gained utility of product Up(X,)and the lost utility of money Um(1)-Um(I-Pi) 2.2.1 Utility of product Figure 1: A concave, bounded, increasing function for Modeling the utility of a product can be traced back to Lan- “1 utility of noney, "approximated with a linear func aster's characteristics theory [15] and Rosen's hedonic price tion for small changes del 24. The hedonic price model es that differenti. The ave assumption can be relaxed when the changes a product can be decomposed into a set of utilities for each that the marginal utility of money is approximately constant. roduct characteristic cording to this model, a product X More formally, we assume that a consumer with income I ith K features can be represented by a K-dimensional vector eceives ney utility Um(n). Paying the price p decrease X=(l,,k), where r* represents the amount or quality small compared to the disposable income 1, the marginal utility product X is then modeled by the function Up(a, .,z") One of the critical issues in this model is how to estimate Under this assumption, the utility of money that the consumer the aggregated utility from the individual product characteris- will lose by paying the price p for product X, can be thereby tics. Based on the hedonic price model, we assume that each represented in a quasi-linear form as follows product characteristic is associated with a weight representing onsumers'desirability towards that characteristic. Under this Um(D)-Um(I-p)=aI-a(I-p)=a(n p,(2) ption, we further refine the definition of overall utility where a(n) denotes the marginal utility for money for someone individual characteristics and an unobserved characteristic E 2.2.3 Challenges Up(X)=Ui(x2…,x)=∑·x+, Given the utility of product and utility of derive the utility surplus as the increase utility, after the purchase. More formally, where B" represents the corresponding weight that the consumer definition for utility surplus is provided as follows assign to the k-th characteristic I. Notice that with Ewe capture the influence of all product characteristics that are not DEFINITION 1: The utility surplus (US), for a consumer explicitly accounted in our model. So, a product that consumers with disposable income I, when buying a product X priced at perceive as high-quality due to a characteristic not explicitly p, is the gain in the utility of product Up minus the loss in the captured in our measurements(e. g. brand name), will end up utility of money Un having a high value of E 2.2.2 Utility of Money US= Up(X)-[Um(D)-Um(I-P)l+E Given the utility of a product, to analyze consumers mo- tivation to trade money for the product, it is also necessary ∑·x+5 to understand the utility of money. This concept is defined as consumers' happiness for owning monetary capital. Based Utility of product on Alfred Marshall's well-established principles[17], utility of Note that s is a product-specific disturbance scalar summarizing money has two basic properties: increasing and concave. unobserved characteristics of product X, and a Increasing: An increase in the amount of money will cause error term that is assumed to be i i.d. across pre an increase in the utility of money. In other words, the nsumers in the selection process and is usually more money someone has, the better. follow a Type I extreme-value distribution. D Concave: The increase in utility, or marginal utility of money, is diminishing as the amount of money increase In this theory model, the key challenge is to estimate the rresponding weights assigned by consumers towards money cample of the utility function for money is shown in Figure 1. Note that with the concave d product dimensions. We discuss this next form of the utility function, the slope is decreasing hence the marginal utility of money is diminishing 100 is more 3. ESTIMATION OF MODEL PARAMETERS rtant for someone with S1000 than for someone with $100.000 In the previous section, we have introduced the background of This also implies that consumers are risk-averse under normal utility theory, characteristics-based theory, and surplus. Recall circumstances. This is because with the same probability to that our main go oal is to identify the best product (with the win or lose, losing N dollars in the assets will cause a drop in highest surplus) for a consumer. This is complicated by the the utility larger than the boost while winning N dollars fact that utility, and therefore surplus, of consumers is private
generates a money utility Um(I). The decision to purchase Xj generates a product utility Up(Xj ) and, simultaneously, paying the price pj decreases the money utility to Um(I − pj ). Assuming that the consumer strives to optimize its own utility, the purchased product Xj is the one that gives the highest increase in utility. This approach naturally generates a ranking order for the products: The products that generate the highest increase in utility should be ranked on top. Thus, to compute the increase in utility, we need the gained utility of product Up(Xj ) and the lost utility of money Um(I) − Um(I − pj ). 2.2.1 Utility of Product Modeling the utility of a product can be traced back to Lancaster’s characteristics theory [15] and Rosen’s hedonic price model [24]. The hedonic price model assumes that differentiated products are described by vectors of objectively measured characteristics. In addition, the utility that a consumer has for a product can be decomposed into a set of utilities for each product characteristic. According to this model, a product X with K features can be represented by a K-dimensional vector X = x1, ... , xK, where xk represents the amount or quality of the k-th characteristic of the product. The overall utility of product X is then modeled by the function Up(x1, ..., xK). One of the critical issues in this model is how to estimate the aggregated utility from the individual product characteristics. Based on the hedonic price model, we assume that each product characteristic is associated with a weight representing consumers’ desirability towards that characteristic. Under this assumption, we further refine the definition of overall utility to be the aggregation of weighted utilities from the observed individual characteristics and an unobserved characteristic ξ: Up(X) = Up(x1 ,...,xK) = K k=1 βk · xk + ξ, (1) where βk represents the corresponding weight that the consumer assign to the k-th characteristic xk. Notice that with ξ we capture the influence of all product characteristics that are not explicitly accounted in our model. So, a product that consumers perceive as high-quality due to a characteristic not explicitly captured in our measurements (e.g. brand name), will end up having a high value of ξ. 2.2.2 Utility of Money Given the utility of a product, to analyze consumers’ motivation to trade money for the product, it is also necessary to understand the utility of money. This concept is defined as consumers’ happiness for owning monetary capital. Based on Alfred Marshall’s well-established principles [17], utility of money has two basic properties: increasing and concave. • Increasing: An increase in the amount of money will cause an increase in the utility of money. In other words, the more money someone has, the better. • Concave: The increase in utility, or marginal utility of money, is diminishing as the amount of money increases. Based on these properties, an example of the utility function for money is shown in Figure 1. Note that with the concave form of the utility function, the slope is decreasing hence the marginal utility of money is diminishing. So, $100 is more important for someone with $1000 than for someone with $100,000. This also implies that consumers are risk-averse under normal circumstances. This is because with the same probability to win or lose, losing N dollars in the assets will cause a drop in the utility larger than the boost while winning N dollars. Figure 1: A concave, bounded, increasing function for “1utility of money,” approximated with a linear function for small changes The concave assumption can be relaxed when the changes in money are small. For most transactions, we often assume that the marginal utility of money is approximately constant. More formally, we assume that a consumer with income I receives a money utility Um(I). Paying the price p decreases the money utility to Um(I − p). Assuming that p is relatively small compared to the disposable income I, the marginal utility of money remains mostly constant in the interval [I − p, I] [17]. Under this assumption, the utility of money that the consumer will lose by paying the price p for product X, can be thereby represented in a quasi-linear form as follows: Um(I) − Um(I − p) = α · I − α · (I − p) = α(I) · p, (2) where α(I) denotes the marginal utility for money for someone with disposable income I. 2.2.3 Challenges Given the utility of product and utility of money, we can now derive the utility surplus as the increase in utility, or excess utility, after the purchase. More formally, the mathematical definition for utility surplus is provided as follows. Definition 1.: The utility surplus (US), for a consumer with disposable income I, when buying a product X priced at p, is the gain in the utility of product Up minus the loss in the utility of money Um. US = Up(X) − [Um(I) − Um(I − p)] + ε i j = k βk · xk + ξ Utility of product − α · p Utility of money + ε Stochastic error (3) Note that ξ is a product-specific disturbance scalar summarizing unobserved characteristics of product X, and ε is a stochastic error term that is assumed to be i.i.d. across products and consumers in the selection process and is usually assumed to follow a Type I extreme-value distribution. ✷ In this theory model, the key challenge is to estimate the corresponding weights assigned by consumers towards money and product dimensions. We discuss this next. 3. ESTIMATION OF MODEL PARAMETERS In the previous section, we have introduced the background of utility theory, characteristics-based theory, and surplus. Recall that our main goal is to identify the best product (with the highest surplus) for a consumer. This is complicated by the fact that utility, and therefore surplus, of consumers is private WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 329
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India and not directly observable. As a result, there exists no"true for establishing the connection between logistic regression and bserved utility" that we can compare with our"model predicte models of discrete user choice utility. "Instead we need to observe the behavior of consumers 3.1.2 Estimation Methodolog and estimate the values of these latent parameters that explain Given Equation 6, we can estimate consumer preferences best the consumer behavior. Furthermore, since we cannot (expressed by the parameters a and P), by observing market assume that we can observe in detail the behavior of individual shares of the different products. One challenge is that we need onsumers, nor can we explicitly ask each consumer for their to know the"demand"for the"buy nothing "option in order to personal"tastes"(e. g, choice of a product, "weight"assigned to product feature, etc. ) we need to extract utilities and derive stimate properly the value P(choicei) in Equation 6. Specifically, we set P(choice)= dibs/dtotal, where dobs is individual preferences by using aggregate data. the observed demand for product j and dtotal is"total demand. The basic idea is the following: If we know the utilities of which includes the demand for the buy-nothing option Taking different products for a consumer, we can estimate the demand for different products, as consumers will behave according to logs in Equation 6 and solving the system 5 heir utility-encoded preferences. So, if we observe the de. mand for various products, we can infer the preferences of the ln()=-ap+∑+6 consumer population for different product aspects. Observing roduct demand is easier than it sounds: For example, we can Such a model can be easily solved to acquire the parameters bserve salesrank on amazon. com and transform salesrank to B and a using any linear regression method, such as ordinar 8, or we can directly observe the transactions at mar- such as eBay and Amazon [11, or we can simply get ExAMPLE 2. Assume that we have a hotel market in New nymous transactions from a merchant In this section, we discuss how we estimate the parameters York, with two hotels: Hotel M(Mandarin Oriental, 5-star), using aggregate demand data. First in Section 3. 1 we discuss and Hotel D(Doubletree, 3-star). From day 1 to 3, we observ that the price for Mandarin Oriental is $500, $480 and $530 pe case where consumers are homogeneous and have night. We also observe a corresponding demand of 400, 470, and similar preferences. Then in Section 3.2 we analyze the more realistic case where consumers have different preferences, which 320 bookings, respectively. Meanwhile, the price for Doubletree is $250, $270 and $225 per night, and its corresponding demand 600, 530 and 680 bookings. Using ou we can 3.1 Homogeneous Consumers: Logit model down the regression equations: 3.1.1 Model specification In(bookings )=-a price+B. stars + hotel +E ( 8) The basic Logit model, introduced by McFadden [18, 19 Here, we divide the unobservable f into a fired effect f that is common for the same hotel (effectively a dummy bina product characteristics. In other words, the weights B and a variable), and an i.i.d. random error term E. Using OLS,we are common across all consumers. Thus, following Definition 1, get a=0.0067 and B=0.64 which express the sensitivity of the the utility surplus for consumer i and product j is written as: consumers to price and their preference for"stars, respectively US;=Vi(o, B)+ej Of course, the assumption of homogeneity of cor where V(a,B)=∑kB,+5-a·p1. Notice that erences is only an ideal case. In reality, consumers are different separate preferences towards product j, captured by Vi(a, B) and their tastes vary. Next, we examine the case where the from non-deterministic aspects of individual consumer behavi consumer have heterogeneous captured by the error term ej 3.2 Heterogeneous Consumers: BLP Model According to the assumption of consumer rationality for utility heterogeneous. In prin maximization, the consumer chooses the product that maximizes ciple, we could observe a customer for a long period of time and utility surplus. Note that the choice is stochastic, given the then use the Logit scheme described above to extract the pref. rror term Ei. Therefore, in our scenario, the probability that erences of each customer. Unfortunatel individual behavior over long periods of time, so it is difficult P(choicei)= P(US,>USi) to estimate the individual preferences for each consumer. (Vl in the same market, 1+1).(5) erences are a function of consumer demographics and purchase Solving this equation, we have [18, 19 context. For example, everything else being equal, honeymoon ers may appreciate a hotel in a romantic remote setting, while P(choicei)= exp(Vi(a, B) business travelers may appreciate more a location with easy 1+∑:exp(V(a,B) access to public transportation. We can therefore characterize In the homogeneous case, all consul ave the each customer by a set of demographic characteristics(e.g, and this age, gender, travel purpose, etc. and make the preference project j(the consumer-specific error term Ei has disappeared Notice that the problem of estimating preferences can be now In this case, the overall preference distribution of the whole population is a mizture of preference distribution of the various expressed as a logistic regression problem. What is worthwhile in this to mention is that this solution is not an adhoc choice, but is consumer types in the population. The main not the a direct derivation from a theory-driven user behavior model setting is that we only observe overall deman demand from each separate consumer group Daniel McFadden got the Nobel prize in Economics in 2000 as, the absolute the parameter estimation and
and not directly observable. As a result, there exists no “true observed utility” that we can compare with our “model predicted utility.” Instead we need to observe the behavior of consumers and estimate the values of these latent parameters that explain best the consumer behavior. Furthermore, since we cannot assume that we can observe in detail the behavior of individual consumers, nor can we explicitly ask each consumer for their personal “tastes” (e.g., choice of a product, “weight” assigned to a product feature, etc.), we need to extract utilities and derive individual preferences by using aggregate data. The basic idea is the following: If we know the utilities of different products for a consumer, we can estimate the demand for different products, as consumers will behave according to their utility-encoded preferences. So, if we observe the demand for various products, we can infer the preferences of the consumer population for different product aspects. Observing product demand is easier than it sounds: For example, we can observe salesrank on Amazon.com and transform salesrank to demand [8], or we can directly observe the transactions at marketplaces such as eBay and Amazon [11], or we can simply get directly anonymous transactions from a merchant. In this section, we discuss how we estimate the parameters using aggregate demand data. First, in Section 3.1 we discuss the simpler case where consumers are homogeneous and have similar preferences. Then in Section 3.2 we analyze the more realistic case where consumers have different preferences, which depend on their demographics and purchase context. 3.1 Homogeneous Consumers: Logit Model 3.1.1 Model Specification The basic Logit model, introduced by McFadden [18, 19], assumes that consumers have “homogeneous preferences” towards product characteristics. In other words, the weights β and α are common across all consumers. Thus, following Definition 1, the utility surplus for consumer i and product j is written as: USi j = Vj (α, β) + ε i j . (4) where Vj (α, β) = k βk · xk j + ξj − α · pj . Notice that we separate preferences towards product j, captured by Vj (α, β), from non-deterministic aspects of individual consumer behavior, captured by the error term εi j . According to the assumption of consumer rationality for utility maximization, the consumer chooses the product that maximizes utility surplus. Note that the choice is stochastic, given the error term εi j . Therefore, in our scenario, the probability that a consumer i chooses product j is: P(choicei j ) = P(USi j > USi l ) (∀l in the same market, l = j). (5) Solving this equation, we have [18, 19]: P(choicei j ) = exp (Vj (α, β)) 1 + l exp (Vl(α, β)). (6) In the homogeneous case, all consumers have the same α and β and this probability is proportional to the market share3 of project j (the consumer-specific error term εi j has disappeared). Notice that the problem of estimating preferences can be now expressed as a logistic regression problem. What is worthwhile to mention is that this solution is not an adhoc choice, but is a direct derivation from a theory-driven user behavior model. Daniel McFadden got the Nobel prize in Economics in 2000 3Market share is defined as the percentage of total sales volume in a market captured by a brand, product, or firm. for establishing the connection between logistic regression and models of discrete user choice. 3.1.2 Estimation Methodology Given Equation 6, we can estimate consumer preferences (expressed by the parameters α and β), by observing market shares of the different products. One challenge is that we need to know the “demand” for the “buy nothing” option in order to estimate properly the value P(choicej ) in Equation 6. Specifically, we set P(choicej ) = dobs j /dtotal , where dobs j is the observed demand for product j and dtotal is “total demand,” which includes the demand for the buy-nothing option.4 Taking logs in Equation 6 and solving the system [5]: ln(dobs j ) = −α · pj + k βk · xk j + ξj . (7) Such a model can be easily solved to acquire the parameters β and α using any linear regression method, such as ordinary least squares (OLS). Example 2. Assume that we have a hotel market in New York, with two hotels: Hotel M (Mandarin Oriental, 5-star), and Hotel D (Doubletree, 3-star). From day 1 to 3, we observe that the price for Mandarin Oriental is $500, $480 and $530 per night. We also observe a corresponding demand of 400, 470, and 320 bookings, respectively. Meanwhile, the price for Doubletree is $250, $270 and $225 per night, and its corresponding demand is 600, 530 and 680 bookings. Using our model, we can write down the regression equations: ln(bookings) = −α · price + β · stars + fhotel + (8) Here, we divide the unobservable ξ into a fixed effect f that is common for the same hotel (effectively a dummy binary variable), and an i.i.d. random error term . Using OLS, we get α = 0.0067 and β = 0.64 which express the sensitivity of the consumers to price and their preference for “stars,” respectively. Of course, the assumption of homogeneity of consumer preferences is only an ideal case. In reality, consumers are different and their tastes vary. Next, we examine the case where the consumer have heterogeneous tastes. 3.2 Heterogeneous Consumers: BLP Model In reality, consumers’ preferences are heterogeneous. In principle, we could observe a customer for a long period of time and then use the Logit scheme described above to extract the preferences of each customer. Unfortunately, we can rarely observe individual behavior over long periods of time, so it is difficult to estimate the individual preferences for each consumer. To allow preferences to vary, though, we can assume that preferences are a function of consumer demographics and purchase context. For example, everything else being equal, honeymooners may appreciate a hotel in a romantic remote setting, while business travelers may appreciate more a location with easy access to public transportation. We can therefore characterize each customer by a set of demographic characteristics (e.g., age, gender, travel purpose, etc.) and make the preference coefficients β to be a function of these demographics. In this case, the overall preference distribution of the whole population is a mixture of preference distribution of the various consumer types in the population. The main challenge in this setting is that we only observe overall demand, and not the demand from each separate consumer group. So, the question 4Since dtotal, appears as a constant in across all equations, the absolute value of dtotal and of the “buy nothing” demand d0 is not relevant to the the parameter estimation and can be ignored. WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 330
WwW 2011- Session: E-commerce March 28-April 1, 2011, Hyderabad, India becomes: How can we find the preferences of various consumer demographic and income distributions P(T)and P( types by simply observing the aggregate product demand? exp(6+apn+∑k時r 3.2.1 Model Specification P(chon)=/1+exp(61+a1n+∑p“) dP(r)dP() We solve this issue by monitoring demand for similar products in different cets, for which we know the distribution of We explain next how we compute this integral and how we onsumers. Since the same product will have the same demand extract the parameters that capture the population preferences from a given demographic group, any differences in demand oss markets can be attributed to the different demographic 3.2.2 Estimation Methodology The following simplified example illustrates the intuition behind With the model in hand, now we discuss how we identify th this approach arameters 5, ar and Br. We apply methods simila to those used in [6, 7] and[25. In general, we estimated the ExAMPLE 3. Consider an example where we have two cities parameters by searching the parameter space in an iterative A and B and two types of consumers: business trip travelers and manner, using the following steps family trip travelers. City A is a business destination with 80% 1.Initialize the parameters 8 o)and 8(o)=(afo), e(o )using of the travelers being business travelers and 20% families. City a random choice of values B is mainly a family destination with 10% business travelers 2. Estimate market shares s; given 0 and and 90% family travelers. In city A, we have two hotels: Hilton 3. Estimate most likely mean utility dj given the market (Al) and Doubletree(A2). In city B, we have again two hotels Hilton(Bi) and Doubletree(B2). Hilton hotels(Al, Bi)have 4. Find the best parameters a and B that minimize the a conference center but no pool, and Doubletree hotels(A2, B2) explained remaining error in dj and evaluate the gen have a pool but no conference center. To keep the example eralized method of moments(GMM)objective function simple, we assume that preferences of consumers do not change when they travel in different cities and that prices are the same 5. Use Nelder-Mead Simplex algorithm to update the pa- By observing demand, we see that demand in city A(business rameter values for 0=(aI, Br) and go to Step 2, until destination)is 820 bookings per day for Hilton and 120 bookings minimizing the GmM objective function for Doubletree. In city B(family destination) the demand is 540 bookings per day for Hilton and 460 bookings for Double- We describe the steps in more detail below. tree. Since the hotels are identical in the two cities, the changes Calculating market share sj: To form the market equations in demand must be the result of different traveler demograph- need two things: the right-hand side sabs that can be observed ics, hinting that a conference center is desirable for business from our transaction data. and the left-hand side si. derived from Equation ll. Unfortunately, the integral in Equation 11 is not analytic. To approximate this integral, we proceed as For this paper, to extract consumer preferences, we use the follows: Given the demographic distribution, we generate"a Random-Coefficient Model (6), introduced by Berry, Levinsohn, consumer randomly, with a known demographic and income and and Pakes, and commonly referred to as the BLP model. This therefore known prefere Then, using the standard Logit model extends the basic Logit model by assuming the coefficients model(Equation 6), we generate the choice of the product for B and a in Equation 6 to be demographic-specific. Let T be a his consumer. For example, assume that we have the following vector representing consumer type, which can specify a particular int demographic distribution of travel purpose and age group purchase context, age group, and so on. In the simplest case, ve can have a binary variable for each consumer group. With lge≤45Age>45 the preferences being now demographic-specific, we write the utility surplus for consumer i, of type T, when buying product 30% 40% j, with features(x3,……,), at price p i to be In this case, we have a 40% probability of generating a"sample US=∑:(r)-a()P+6+ consumer" with family travel purpose and age above 45. By (9) repeating the process and obtaining Nr samples of demographic T and NI samples of income I, we can compute an unbiased For the Logit model, in Equation 4, we used V(a, B)to stylis. estimator of the Equation 1l integral: 5 tically separate the population preferences from the idiosyn- ep(+a1+∑k跨r“) cratic behavior of the consumer. We now do the same for the BLP model, separating the mean population preferences from 83(631)~MNr7+2∞p(+a1P+∑kTr) the demographic-specific preferences. So, we write B(T) (Bk+BTr), where Bk is the mean of the preference distribu Estimate utility d Since we know how to compute tion, and Br is a vector capturing the variation in the preferen market shares from the parameters, we can now find a value from different consumer types. Similarly, we model a as a func- of oj that best"fits"the observed market shares.(Notice that, conditional on 8=(or, Br), market share s; can be viewed as ion of income r: a(r)=(a+arr ). Notice that we assume a function of the mean utility 8. )We apply the contraction aI and Br to be independent. We rewrite US;as mapping method recommended by Berry [6], which suggest computing the value for d using an iterative apy ∑(+時r)·+5-(+aP) (1 (91).(13) We use d=-a·p+∑ The procedure is guaranteed to converge [6] and find d, that zi+5, to represent the mean satisfies s, (6, 10)=s9 y the Logit me the choice probability for j, by integrating over the population We use N= Nr= 100 in our study. 331
becomes: How can we find the preferences of various consumer types by simply observing the aggregate product demand? 3.2.1 Model Specification We solve this issue by monitoring demand for similar products in different markets, for which we know the distribution of consumers. Since the same product will have the same demand from a given demographic group, any differences in demand across markets can be attributed to the different demographics. The following simplified example illustrates the intuition behind this approach. Example 3. Consider an example where we have two cities, A and B and two types of consumers: business trip travelers and family trip travelers. City A is a business destination with 80% of the travelers being business travelers and 20% families. City B is mainly a family destination with 10% business travelers and 90% family travelers. In city A, we have two hotels: Hilton (A1) and Doubletree (A2). In city B, we have again two hotels: Hilton (B1) and Doubletree (B2). Hilton hotels (A1, B1) have a conference center but no pool, and Doubletree hotels (A2, B2) have a pool but no conference center. To keep the example simple, we assume that preferences of consumers do not change when they travel in different cities and that prices are the same. By observing demand, we see that demand in city A (business destination) is 820 bookings per day for Hilton and 120 bookings for Doubletree. In city B (family destination) the demand is 540 bookings per day for Hilton and 460 bookings for Doubletree. Since the hotels are identical in the two cities, the changes in demand must be the result of different traveler demographics, hinting that a conference center is desirable for business travelers. For this paper, to extract consumer preferences, we use the Random-Coefficient Model [6], introduced by Berry, Levinsohn, and Pakes, and commonly referred to as the BLP model. This model extends the basic Logit model by assuming the coefficients β and α in Equation 6 to be demographic-specific. Let Ti be a vector representing consumer type, which can specify a particular purchase context, age group, and so on. In the simplest case, we can have a binary variable for each consumer group. With the preferences being now demographic-specific, we write the utility surplus for consumer i, of type Ti , when buying product j, with features x1 j ,...,xk j , at price pj to be: USi j = k βk(Ti ) · xk j − α(Ii ) · pj + ξj + ε i j . (9) For the Logit model, in Equation 4, we used V (α, β) to stylistically separate the population preferences from the idiosyncratic behavior of the consumer. We now do the same for the BLP model, separating the mean population preferences from the demographic-specific preferences. So, we write βk(Ti ) = ¯ βk + βk T Ti , where ¯ βk is the mean of the preference distribution, and βk T is a vector capturing the variation in the preferences from different consumer types. Similarly, we model αi as a function of income Ii : α(Ii ) = α¯ + αI Ii . Notice that we assume αI and βT to be independent. We rewrite USi j as: USi j = k ¯ βk + βk T Ti · xk j + ξj − α¯ + αI Ii · pj + εi j . (10) We use δj = −α¯ · pj + k ¯ βk · xk j + ξj to represent the mean utility of product j. Then, as in the Logit model, we derive the choice probability for j, by integrating over the population demographic and income distributions P(T) and P(I): P (choicej ) = exp δj + αI Iipj + k βk T Tixk j 1 + l exp δl + αI Iipl + k βk T Tixk l dP (T) dP (I) (11) We explain next how we compute this integral and how we extract the parameters that capture the population preferences. 3.2.2 Estimation Methodology With the model in hand, now we discuss how we identify the unknown parameters δj , αI and βT . We apply methods similar to those used in [6, 7] and [25]. In general, we estimated the parameters by searching the parameter space in an iterative manner, using the following steps: 1. Initialize the parameters δ (0) j and θ(0) = (α(0) I , β(0) T ) using a random choice of values. 2. Estimate market shares sj given θ and δ. 3. Estimate most likely mean utility δj given the market shares. 4. Find the best parameters α¯ and ¯ βk that minimize the unexplained remaining error in δj and evaluate the generalized method of moments (GMM) objective function. 5. Use Nelder-Mead Simplex algorithm to update the parameter values for θ = (αI , βT ) and go to Step 2, until minimizing the GMM objective function. We describe the steps in more detail below. Calculating market share sj : To form the market equations (i.e., model predicted market share = observed market share), we need two things: the right-hand side sobs j that can be observed from our transaction data, and the left-hand side sj , derived from Equation 11. Unfortunately, the integral in Equation 11 is not analytic. To approximate this integral, we proceed as follows: Given the demographic distribution, we “generate” a consumer randomly, with a known demographic and income and, therefore, known preferences. Then, using the standard Logit model (Equation 6), we generate the choice of the product for this consumer. For example, assume that we have the following joint demographic distribution of travel purpose and age group: Age ≤ 45 Age > 45 Business 15% 15% F amily 30% 40% . In this case, we have a 40% probability of generating a “sample consumer” with family travel purpose and age above 45. By repeating the process and obtaining NT samples of demographics Ti and NI samples of income Ii , we can compute an unbiased estimator of the Equation 11 integral:5 sj (δj |θ) ∼ 1 NI 1 NT NI Ii NT T i exp δj + αI Iipj + k βk T Tikxk j 1 + l exp δl + αI Iipl + k βk T Tikxk l . (12) Estimate mean utility δj : Since we know how to compute market shares from the parameters, we can now find a value of δj that best “fits” the observed market shares. (Notice that, conditional on θ = (αI , βT ), market share sj can be viewed as a function of the mean utility δj .) We apply the contraction mapping method recommended by Berry [6], which suggests computing the value for δ using an iterative approach: δ (t+1) j = δ (t) j + (ln(s obs j ) − ln(sj (δ (t) j |θ))). (13) The procedure is guaranteed to converge [6] and find δj that satisfies sj (δj |θ) = sobs j . 5We use NT = NI = 100 in our study. WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 331
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India Minimize error, evaluate GMM objective function: Once These models use the concept of surplus mainly as a conceptual we have the market shares and the mean utility parameters, we tool to infer consumer preferences towards different product need to find the most likely demographic-specific weight devia- characteristics In our work, the concept of surplus is directly ions 0=(a1, Br). Of course, different values for 8=(a1, Br) used to find the product that is the best value for money"for will lead to different mean utilities and market shares. Heno a given consumer. This simple idea is at the core of our work we need to find a criterion for identifying the best solution. We an ill demonstrate in the experimental evaluation, it perform this in two steps: First, we use Instrumental Variables can lead to significant improvements in the quality of product (v)[13] to estimate the mean weights a and B, and extract search results the unobserved error term E from the mean utility function Surplus-based Ranking: The first approach is to use th (0)=6(0)-C∑.x2-ap estimated surplus for each product and rank the available prod (14) ucts in decreasing order of surplus. Therefore, at the top will have the products that are the best value"for consum In our study, we use the average price of the "same-star rating" for a given price. We define Consumer Surplus for consumer i hotels in other markets as the instrument for price of a particular from product j as the "normalized utility surplus, "the surplus hotel to ensure that we do not have a correlation of the error term US divided by the mea with a variable in our regression. Then, using the generalized method of moments, we base our analysis on the moment condition that the mean of the unobserved error term E is CS,=NormalizedUS,=>=US.(16) uncorrelated with the instrumental variable /v. Thus in our case we are trying to minimize the objection function: In the general, non-personalized case, if we were ranking prod- ucts based on the"training"demand data then, in theory, our GMMobj(6)=E(6)·IV] (15) product ranking would be similar to a "best selling" ranking The products that generate that largest surplus are the ones Once we identify the mean utility for a given set of weight devi. that would also generate the highest sales. (Notice that ratio- ations 0=(a1, Br), we note the value of the GMM objective nal consumers prefer the products that generate the highest Inction GMMobj (0). Then, we use the Nelder- Mead Simplex plus. )However, when ranking products that are available lgorithm [21] to search for the optimal 8=(ai, Br) that today, the surplus-based ranking may be different for a variety izes the GMM objective function. This whole process of reasons: the product p eventually identifies the heterogeneous weights that different products a better "value for money, we may have a new product consumers assign to product price, a(r)=a+aiI', and those being assigned to product characteristics, B(T)=B+Br.T time-dependent (e. g, the value of being next to a lake may be ositive during warm weather and negative during the winter) ExAMPLE 4. To illustrate this better, let's again look at ET- Personalized Surplus-based Ranking: In Section 3.2 we ample 3, We know that, for business traveler, the utility sur. escribed how to estimate the value that consumers place on plus from hotel Ai(conference center, no pool) is US(A1) different product features, based on their own demographics and SA1+(Bcon.1+Pool 0)+e, and for family travelers, the corre- purchase context. The main outcome is that the value(surplus) sponding utility surplus is USP(A1)=0A+(Bcon/1+BBool-0)+6. that consumers get from a particular purchase is different than By B. we denote the deviations from the population mean for the average surplus for the overall population. This means that center"and "pool"and even the best possible ranking for the general population may not be optimal for an individual consumer. by B. we denote the respective deviations for family traveler Similarly, we can write down the utilities for hotels A2, Bi and Therefore, we extend our ranking to include a personalizatio B2. Following the estimation steps discussed above, we discover ponent. To compute the personalized surplus, ey have the same preferences regan e =0.5. In other words, tics and purchase context(e.g,35-49 years old,male,S100K at family travelers have B the consumer to give the appropriate demographic characteris- nd conference center. On the other hand, for business travelers, their prefer. viation matrices Br and al. It is then easy to compute the ence towards"conference center"is much higher than towards personalized "value for money"for this individual consumer, pool, "with Conf=0.9 and Pool =0.1, respectively and rank products accordingly. Notice that the consumer has Next, we explain how we leverage the above knowledge for the incentive to reveal demographics in this scenario. building a better ranking model for product search ExAMPLE 5. For better understanding, let's re-consider the 4. RANKING USING UTILITY SURPLUS previous setting of the two hotels Al and A2 for city A from ET- amples 3 Suppose that two consumers are traveling to city So far, we have described models for inferring the preferences A on the same day: C1, a 35-49 years old business traveler of consumers using a utility model and aggregate demand data. with an income $50, 000-100, 000, and C2, a 25-34 years old family traveler, with an income less than $50,000 Due to space restrictions, we do not describe the GMM method in detail Since these tuo travelers belong to different demographic groups This approach is typically better than just following the function gradi- and travel with different purposes, their preferences towards /programming-doc/gsl/gal-ref they obtain from Al and A2 varies In our application, the computational time for each call (i.e, the ness traveler gets higher utility from Al due to the specialized inner loop)to the GMM objective function to solve for the mean conference center services, whereas the family traveler find A2 tility is around 3 minutes on average. The e, the outer loop) by minimizing the GMM objective function takes more valuable due to the pool and price. This personalization 4a verage of 20 calls. The total time for the estimation is around component allows each consumer to identify the product that D-60 mi the“ best value for the money
Minimize error, evaluate GMM objective function: Once we have the market shares and the mean utility parameters, we need to find the most likely demographic-specific weight deviations θ = (αI , βT ). Of course, different values for θ = (αI , βT ) will lead to different mean utilities and market shares. Hence, we need to find a criterion for identifying the best solution. We perform this in two steps: First, we use Instrumental Variables (IV) [13] to estimate the mean weights α¯ and β¯, and extract the unobserved error term ξ from the mean utility function: ξ(θ) = δ(θ) − ( k ¯ βk · xk − α¯ · p). (14) In our study, we use the average price of the “same-star rating” hotels in other markets as the instrument for price of a particular hotel to ensure that we do not have a correlation of the error term with a variable in our regression. Then, using the generalized method of moments, 6 we base our analysis on the moment condition that the mean of the unobserved error term ξ is uncorrelated with the instrumental variable IV . Thus, in our case we are trying to minimize the objection function: GMMobj(θ) = E[ξ (θ) · IV ]. (15) Iterate until GMM objective function is minimized: Once we identify the mean utility for a given set of weight deviations θ = (αI , βT ), we note the value of the GMM objective function GMMobj(θ). Then, we use the Nelder-Mead Simplex algorithm [21] to search for the optimal θ∗ = (α∗ I , β∗ T ) that minimizes the GMM objective function.7 This whole process eventually8 identifies the heterogeneous weights that different consumers assign to product price, α(Ii ) = α¯∗+α∗ I ·Ii , and those being assigned to product characteristics, β(Ti ) = β¯∗ + β∗ T · Ti . Example 4. To illustrate this better, let’s again look at Example 3, We know that, for business traveler, the utility surplus from hotel A1 (conference center, no pool) is USB(A1) = δA1 + (βB conf · 1+βB pool · 0)+, and for family travelers, the corresponding utility surplus is USF (A1) = δA1+(βF conf ·1+βF pool·0)+. By βB • we denote the deviations from the population mean for business travelers towards “conference center” and “pool” and by βF • we denote the respective deviations for family travelers. Similarly, we can write down the utilities for hotels A2, B1 and B2. Following the estimation steps discussed above, we discover that family travelers have βF conf = βF pool = 0.5. In other words, they have the same preferences regarding a pool and conference center. On the other hand, for business travelers, their preference towards “conference center” is much higher than towards “pool,” with βP conf = 0.9 and βF pool = 0.1, respectively. Next, we explain how we leverage the above knowledge for building a better ranking model for product search. 4. RANKING USING UTILITY SURPLUS So far, we have described models for inferring the preferences of consumers using a utility model and aggregate demand data. 6Due to space restrictions, we do not describe the GMM method in detail here. We refer the interested reader to [12] for further explanations. 7This approach is typically better than just following the function gradient. See http://linux.math.tifr.res.in/programming-doc/gsl/gsl-ref_ 34.html for an open source implementation and details. 8In our application, the computational time for each call (i.e., the inner loop) to the GMM objective function to solve for the mean utility is around 3 minutes on average. The global parameter search (i.e., the outer loop) by minimizing the GMM objective function takes an average of 20 calls. The total time for the estimation is around 40-60 minutes. These models use the concept of surplus mainly as a conceptual tool to infer consumer preferences towards different product characteristics. In our work, the concept of surplus is directly used to find the product that is the “best value for money” for a given consumer. This simple idea is at the core of our work and as we will demonstrate in the experimental evaluation, it can lead to significant improvements in the quality of product search results. Surplus-based Ranking: The first approach is to use the estimated surplus for each product and rank the available products in decreasing order of surplus. Therefore, at the top we will have the products that are the “best value” for consumers, for a given price. We define Consumer Surplus for consumer i from product j as the “normalized utility surplus,” the surplus US¯ (i) j divided by the mean marginal utility of money ¯α. CSj = Normalized USj = t 1 α¯ US¯ (i) j . (16) In the general, non-personalized case, if we were ranking products based on the “training” demand data then, in theory, our product ranking would be similar to a “best selling” ranking: The products that generate that largest surplus are the ones that would also generate the highest sales. (Notice that rational consumers prefer the products that generate the highest surplus.) However, when ranking products that are available today, the surplus-based ranking may be different for a variety of reasons: the product price may have changed, making some products a better “value for money,” we may have a new product in the market, or the value of some product features may be time-dependent (e.g., the value of being next to a lake may be positive during warm weather and negative during the winter). Personalized Surplus-based Ranking: In Section 3.2 we described how to estimate the value that consumers place on different product features, based on their own demographics and purchase context. The main outcome is that the value (surplus) that consumers get from a particular purchase is different than the average surplus for the overall population. This means that even the best possible ranking for the general population may not be optimal for an individual consumer. Therefore, we extend our ranking to include a personalization component. To compute the personalized surplus, we can ask the consumer to give the appropriate demographic characteristics and purchase context (e.g., 35-49 years old, male, $100K income, business traveler) and then use the corresponding deviation matrices βT and αI . It is then easy to compute the personalized “value for money” for this individual consumer, and rank products accordingly. Notice that the consumer has the incentive to reveal demographics in this scenario. Example 5. For better understanding, let’s re-consider the previous setting of the two hotels A1 and A2 for city A from Examples 3 and 4. Suppose that two consumers are traveling to city A on the same day: C1, a 35-49 years old business traveler, with an income $50,000-100,000, and C2, a 25-34 years old family traveler, with an income less than $50,000. Since these two travelers belong to different demographic groups and travel with different purposes, their preferences towards “conference center” and “pool” are different. Thus, the surplus they obtain from A1 and A2 varies. For example, the business traveler gets higher utility from A1 due to the specialized conference center services, whereas the family traveler find A2 more valuable due to the pool and price. This personalization component allows each consumer to identify the product that is the “best value for the money.” WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 332
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India 5. EXPERIMENTAL SETUP Variable Coef. Variabl Coef For our experimental evaluation, we instantiated our 0.1157 of competitors -0.0930 framework using as target application the area of hotel Demand data: Travelocity, a large hotel booking 013 of int. amenities rovided us with the set of all hotel booking transactions for Readability(SMOG) 2117 randomly selected hotels over the United States. The transactions covered the period from November 2008 until Jan- 1020 uary 2009. Based on the given transactions, we were able to Hotel CI ompute the market shares of each hotel in each local market Review Rating (i.e, metropolitan area), for each day. Consumer demographics data: To measure the demo- eview service Review Checkin -O1151 Review- Bus Service 0.143 graphics of consumers in each target market, we used data rovided by Trip Advisor: The consumers that write revie Table 1: Estimation results for mean weights (listing about hotels on TripAdvisor also identify their travel purpose only statistically significant coefficients, with P< 0.05) 17,18-24,25-34,3549,50-64,65+).° Based on the%(13 (business, romance, family, friend, other) and age grou were able to identify the types of travelers for each destination. ection, we present our findings from the empirical estimation. To ensure the quality of the data, we computed the Jensen- in Section 6.1 we present the results on estimating the Shannon divergence of the demographic distribution extracted parameters, which correspond to the consumer prefer- hen, in Section 6.2, we show that our models generate from Trip Advisor with the corresponding traveler information from Travelocity, whenever available. The distributions were gnificantly better rankings than the existing baseline very similar with an average divergence of just 0.03 Hotel location characteristics: We used geo-mapping 6.1 Interpretation of Estimated Weights search tools(in particular the Bing Maps API)and social ge We present only the estimation results for the BLP model, as tags( from geonames. org) to identify the "external amenities" the results are strongly superior and closer to reality compared around the hotel. We also used image classification together Results for general population: Table I shows th with Mechanical Turk to examine whether there is a nearby mation results for the weights, a and B, for the d beach, a nearby lake, a downtown area, and whether the hotel ariables in our model that have a statistically significant is close to a highway [16]. We extracted these characteristics on demand. From the results, "beach"presents the highest within an area of 0.25-mile, 0.5 mile, l-mile, and 2-mile radius Hotel service characteristics: We extracted the service We also found significant impacts from service characteristics haracteristics from the reviews from TripAdvisor. Each review and quality characteristics of word-of-mouth. Meanwhile, "price" provides a general rating of the hotel, plus provides seven in- presents a positive sign, which is consistent with the"law of demand"in reality and indicates that the higher the price, the oom. Location, Cleanliness. Service. Check-in, and Business lower the demand. The negative sign on subjectivity means ce. We used the average ratings of each hotel across these that customers are positive influenced by reviews that describe seven characteristic ether with the general review rati factual characteristics of hotels, and do not want to read per- We also used the hotel description information from Travelocity sonal stories of reviewers.(Notice that this is independent of Orbitz, and Expedia, to identify the "internal amenities"(e.g. the review polarity. Notable is the negative sign on the re pool, spa, etc. vieuvalue rating, which indicates that hotels that receive Stylistic characteristics of online reviews: Finally, we "high value"rating have lower demand. This is not surprising extracted indicators that measure not the polarity of the reviews these are the hotels that are "undiscovered "and therefore h but rather some stylistic characteristics of the available revier lower demand and prices than otherwise expected We examined 2 text-style features: subjectivity and"readabil- Results for specific demographics: We also obtained ity "of reviews [10. Also, since prior research suggested that the demographic-specific deviations from the mean. We used disclosure of identity information is associated with changes in ase contert and age group as the demographic dimensions subsequent online product sales 9, we measured the percentage for our experiments(see Section 5) of reviewers for each hotel who reveal their real name or location The value of demographic-specific deviation shows the"sen- information on their profile web pages sitivity of evaluation"for a product characteristic, within Figure 2 shows the histograms of the important continuous particular consumer type. For example, customers on a roman variables, together with their correlations and scatterplots that tic trip are sensitive to hotel characteristics like"class"o illustrate the joint distributions of the variable pairs. "close to a beach", but they are less interested to know whether not the hotel is close to highway exits. On the contrary, 6. EXPERIMENTAL RESULTS customers on a business trip are more sensitive to hotel char- acteristics like"internal amenities"or" easy access to highway In the previous section, we have discussed how we retrieved hereas they are likely influenced by the star rating of the hotel, ferent hotel characteristics through various sources. In this to romance travelers. Figure 3 shows in details the aluation deviations We also examined age-specific preferences. Again, we found "traveler residence location, ""traveling with,""usual travel style"and strong evidence for the deviation of weights associated with "usual travel purpose. " However, these fields were relatively sparsely erefore. we did not use these variables for our stud different age groups. Especially for"reviewer overall rating"and We have also extracted service-related variables by mining directly review count", the deviations become quite striking. Figure 4 the text of the reviews 2, 23 but the additional information did not hows that customers from age 18 to 34 tend to be more sensitive improve our model in a statistically significant manner to online reviews, compared to older ages. In particular, they
5. EXPERIMENTAL SETUP For our experimental evaluation, we instantiated our model framework using as target application the area of hotel search. Demand data: Travelocity, a large hotel booking system, provided us with the set of all hotel booking transactions, for 2117 randomly selected hotels over the United States. The transactions covered the period from November 2008 until January 2009. Based on the given transactions, we were able to compute the market shares of each hotel in each local market (i.e., metropolitan area), for each day. Consumer demographics data: To measure the demographics of consumers in each target market, we used data provided by TripAdvisor: The consumers that write reviews about hotels on TripAdvisor also identify their travel purpose (business, romance, family, friend, other) and age group (13- 17, 18-24, 25-34, 35-49, 50-64, 65+).9 Based on the data, we were able to identify the types of travelers for each destination. To ensure the quality of the data, we computed the JensenShannon divergence of the demographic distribution extracted from TripAdvisor with the corresponding traveler information from Travelocity, whenever available. The distributions were very similar with an average divergence of just 0.03. Hotel location characteristics: We used geo-mapping search tools (in particular the Bing Maps API) and social geotags (from geonames.org) to identify the “external amenities” (e.g., shops, bars, etc) and public transportation in the area around the hotel. We also used image classification together with Mechanical Turk to examine whether there is a nearby beach, a nearby lake, a downtown area, and whether the hotel is close to a highway [16]. We extracted these characteristics within an area of 0.25-mile, 0.5 mile, 1-mile, and 2-mile radius. Hotel service characteristics: We extracted the service characteristics from the reviews from TripAdvisor. Each review provides a general rating of the hotel, plus provides seven individual ratings on the following service characteristics: Value, Room, Location, Cleanliness, Service, Check-in, and Business Service. We used the average ratings of each hotel across these seven characteristics, together with the general review rating.10 We also used the hotel description information from Travelocity, Orbitz, and Expedia, to identify the “internal amenities” (e.g., pool, spa, etc.) Stylistic characteristics of online reviews: Finally, we extracted indicators that measure not the polarity of the reviews but rather some stylistic characteristics of the available reviews. We examined 2 text-style features: “subjectivity” and “readability” of reviews [10]. Also, since prior research suggested that disclosure of identity information is associated with changes in subsequent online product sales [9], we measured the percentage of reviewers for each hotel who reveal their real name or location information on their profile web pages. Figure 2 shows the histograms of the important continuous variables, together with their correlations and scatterplots that illustrate the joint distributions of the variable pairs. 6. EXPERIMENTAL RESULTS In the previous section, we have discussed how we retrieved different hotel characteristics through various sources. In this 9There are other demographics available as well, such as “gender,” “traveler residence location,”“traveling with,”“usual travel style” and “usual travel purpose.” However, these fields were relatively sparsely populated. Therefore, we did not use these variables for our study. 10We have also extracted service-related variables by mining directly the text of the reviews [2, 23] but the additional information did not improve our model in a statistically significant manner. Variable Coef. Variable Coef. Price 0.1157 # of competitors -0.0930 Avg review length 0.0291 Crime -0.4226 # of ext. amenities 0.0013 # of int. amenities 0.0048 Readability (SMOG) 0.2308 Beach 0.5498 Spelling errors -0.0764 Lake -0.1884 Avg. Subjectivity -1.3468 Transport 0.00005 Dev. Subjectivity -2.9106 Highway 0.2082 % of non-anon reviews 0.0892 Downtown 0.0161 Hotel Class 0.0317 # of reviews -0.3897 Review Rating 0.0835 Review Value -0.2988 # of rooms 0.1525 Review Location 0.0845 Review Clean 0.1309 Review Service 0.0105 Review Checkin -0.1151 Review Bus Service 0.1432 Table 1: Estimation results for mean weights (listing only statistically significant coefficients, with p < 0.05) section, we present our findings from the empirical estimation. First, in Section 6.1 we present the results on estimating the model parameters, which correspond to the consumer preferences. Then, in Section 6.2, we show that our models generate significantly better rankings than the existing baselines. 6.1 Interpretation of Estimated Weights We present only the estimation results for the BLP model, as the results are strongly superior and closer to reality compared to the results from the Logit model. Results for general population: Table 1 shows the estimation results for the mean weights, α¯ and β¯, for the different variables in our model that have a statistically significant effect on demand. From the results, “beach” presents the highest positive impact compared to the other location characteristics. We also found significant impacts from service characteristics and quality characteristics of word-of-mouth. Meanwhile, “price” presents a positive sign, which is consistent with the “law of demand” in reality and indicates that the higher the price, the lower the demand. The negative sign on subjectivity means that customers are positive influenced by reviews that describe factual characteristics of hotels, and do not want to read personal stories of reviewers. (Notice that this is independent of the review polarity.) Notable is the negative sign on the review value rating, which indicates that hotels that receive a “high value” rating have lower demand. This is not surprising: these are the hotels that are “undiscovered” and therefore have lower demand and prices than otherwise expected. Results for specific demographics: We also obtained the demographic-specific deviations from the mean. We used purchase context and age group as the demographic dimensions for our experiments (see Section 5). The value of demographic-specific deviation shows the “sensitivity of evaluation” for a product characteristic, within a particular consumer type. For example, customers on a romantic trip are more sensitive to hotel characteristics like “class” or “close to a beach”, but they are less interested to know whether or not the hotel is close to highway exits. On the contrary, customers on a business trip are more sensitive to hotel characteristics like “internal amenities” or “easy access to highway”, whereas they are likely influenced by the star rating of the hotel, compared to romance travelers. Figure 3 shows in details the evaluation deviations among different types. We also examined age-specific preferences. Again, we found strong evidence for the deviation of weights associated with different age groups. Especially for “reviewer overall rating” and “review count”, the deviations become quite striking. Figure 4, shows that customers from age 18 to 34 tend to be more sensitive to online reviews, compared to older ages. In particular, they WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 333
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India Hotel Data Correlograms 024 聊[[[[[[[[[[[a 1[a[0[匚a [m□[T8 □[[□[[[[0[[□鳞[ 8□a7[0[08[a1[2[245[3[0[ [a4[at[a[47[4014[0匚 0[08[[[4[4am□am[s 司[s[[a2[04[9[sn [[a3□a1[s2 和聊图和 Figure 2: Diagonal: Histograms for the continuous variables in the data set (market share, price, ext. amenities, value room, location, checkin, service, business service, rooms, class, crime, int_amenities, review_count, cleaningness, tripadvisor rating). Top: correlations of variable pairs. Bottom: scatterplots with ellipses for joint distributions of variable pairs. Most Price Price Hotel f of of ity Travelocity Booked Low to High to Class Reviews Rooms Amenities Price with Low Class New York 71% San francisco 892903 alt Lake City 69% % Significance (Sign Test Table 2: Consumer-surplus-based ranking vs existing baselines. The percentage shows the number of users than indicated preference for the ranking of our surplus-based ranking, compared to the corresponding baseline ranking. The comparison was a pairwise blind comparison of rankings of 10 hotels. Each cell corresponds to a separate test using N= 200 users, for a total of 12000 comparisons. Across all sons, our ranking fares better in a statistically significant manner. Given the noise inherent in the Mechanical Turk ratings(some users may give andom answers in such tests and vote equally for both approaches), the results show a strong superiority of the surplus-based rankings. are 10-15 times more sensitive to online reviews compared to users, the majority of users preferred our surplus-based rankings. the age demographic of65+ year old. in a statistically significant manner. Given that some MTurk users may be giving us random results, we expect the"real" 6.2 User Study: Ranking Comparison performance of our algorithm to be even better than indicated With the estimated weights, we can derive the consumer the numbers in table 2 Qualitative analysis: We also asked consumers why they chose generate rankings. In our experiments, we used six metropolitan a particular ranking, to better understand how users interpret areas for which we generated hotel rankings( we used big cities, that we have a meaningful number of hotels to rank) indicated that our ranking promotes the idea that price is no Rankings based on the average consumer surplus vs. the only main factor in rating the quality of products. Moreover. users strongly preferred the diversity provided by our rankin existing ranking baselines. We used 10 different, existing across both price and quality. In contrast, the other ranking ankings as a baseline for comp baseline against surplus-based ranking. For each city approaches tend to list products of only one type(e.g,hotels nd each ranking baselin performed pair-wise blind tests, with high review ratings are often very expensive hotels and asking 200 anonymous users on Amazon Mechanical Turkt he"most booked"are often 3-star mediocre hotels). We should compare pairs of rankings and tell us which of the hotel rankin mphasize at this point that our algorithm does not try explicitly lists they prefer better. In all 60 comparisons, each using 200 to introduce diversity in the results. This is a direct outcome gment of the m
Figure 2: Diagonal: Histograms for the continuous variables in the data set (market share, price, ext amenities, value, room, location, checkin, service, business service, rooms, class, crime, int amenities, review count, cleaningness, tripadvisor rating). Top: correlations of variable pairs. Bottom: scatterplots with ellipses for joint distributions of variable pairs. Most Price Price Hotel # of # of # of Diversity: City TripAdvisor Travelocity Booked Low to High to Class Reviews Rooms Amenities Price with High Low Class New York 77% 63% 61% 57% 71% 88% 76% 89% 60% 80% Los Angeles 72% 58% 71% 59% 84% 89% 87% 86% 69% 76% San Francisco 79% 57% 65% 62% 70% 82% 68% 79% 79% 72% Orlando 83% 81% 62% 63% 73% 79% 73% 79% 61% 79% New Orleans 61% 69% 60% 78% 69% 80% 72% 91% 58% 85% Salt Lake City 61% 80% 69% 66% 79% 83% 73% 70% 76% 79% Significance p = 0.05 p = 0.01 p = 0.001 (Sign Test Level ≥ 56% ≥ 59% ≥ 61% N = 200) Table 2: Consumer-surplus-based ranking vs. existing baselines. The percentage shows the number of users than indicated preference for the ranking of our surplus-based ranking, compared to the corresponding baseline ranking. The comparison was a pairwise blind comparison of rankings of 10 hotels. Each cell corresponds to a separate test using N = 200 users, for a total of 12000 comparisons. Across all comparisons, our ranking fares better in a statistically significant manner. Given the noise inherent in the Mechanical Turk ratings (some users may give random answers in such tests and vote equally for both approaches), the results show a strong superiority of the surplus-based rankings. are 10-15 times more sensitive to online reviews compared to the age demographic of “65+ year old.” 6.2 User Study: Ranking Comparison With the estimated weights, we can derive the consumer surplus for each product (hotel), which can then be used to generate rankings. In our experiments, we used six metropolitan areas for which we generated hotel rankings (we used big cities, so that we have a meaningful number of hotels to rank). Rankings based on the average consumer surplus vs. existing ranking baselines.: We used 10 different, existing rankings as a baseline for comparison, and we compared each baseline against our own surplus-based ranking. For each city and each ranking baseline, we performed pair-wise blind tests, asking 200 anonymous users on Amazon Mechanical Turk11 to compare pairs of rankings and tell us which of the hotel ranking lists they prefer better. In all 60 comparisons, each using 200 11We restricted participation to US-based users. Mechanical Turk users are representative of the general US population. users, the majority of users preferred our surplus-based rankings, in a statistically significant manner. Given that some MTurk users may be giving us random results, we expect the “real” performance of our algorithm to be even better than indicated in the numbers in Table 2. Qualitative analysis: We also asked consumers why they chose a particular ranking, to better understand how users interpret the surplus-based ranking. Many users that liked our ranking indicated that our ranking promotes the idea that price is not the only main factor in rating the quality of products. Moreover, users strongly preferred the diversity provided by our ranking across both price and quality. In contrast, the other ranking approaches tend to list products of only one type (e.g., hotels with high review ratings are often very expensive hotels and the “most booked” are often 3-star mediocre hotels). We should emphasize at this point that our algorithm does not try explicitly to introduce diversity in the results. This is a direct outcome of our economic-based approach: If a segment of the market is systematically underpriced (hence making the “best deals” a homogeneous list), then we expect the market forces to fix this WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 334
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India ROMANCE BUSINESS FAMILY FRIENDS a Figure 3: Sensitivity towards different hotel characteristics for Romance, Business, Family, and Friends-getawa travelers. We present absolute values, without polarity, to illustrate sensitivity on different product characteristics. Personalized rankings vs rankings using average con sumer surplus: We generated a few personalized rankings for different cities based on consumer-specific attributes, such as travel purpose. We conducted blind comparisons in a pair-wise fashion, based on 200 anonymous AMT users, for each compar son. Since we did not know the demographics of the users, we asked the MTurk workers to"select the hotel ranking you would refer to use while trying to help a business traveler to book a hotel in New york "12 Based on the user responses, customers strongly preferred the 131718242534349506465+ Other personalized ranking that was tailored for a particular travel purpose using our technique (p=0.01, sign test). For example, in our NYC experiment, 77% customers indicated their prefer Figure 4: Sensitivity of different age groups to review rating &z review count: Ages groups 18-24 and 25- ces towards the business-oriented ranking (ranking tailored 34 pay more attention to online reviews compared to for business travelers) rather than the average-level other age groups. (ran 81% customers did so towards the family-oriented nking tailored for family trip travelers). The resu City Business Romand amily Friends Qualitative analysis: When we asked users'opinions for these New York omparisons, users did not bring up the issue of diversity. This was expected, as even our non-personalized rankings were al- 3 ready diverse. Instead, we found customers considering the context and expectations for a given trip. For example, users Table 3: Personalized surplus-based rankings indicated that hotels for business trips do not necessarily need to be luxury, but need to provide a quiet business environment personalized, surplus-based rankings. The perce shows the number of users than indicated pref LfntnBe and easy access to highway or public transportation. On the for the personalized ranking compared to the general other hand, for romantic trips, users strongly preferred the luxury services with higher non-personalized surplus-based ranking. Each cell of These results highly dovetail with our empirical estimation suggesting 3200 comparisons. In all cases, the personalized ap- eal purchase motivation behind the scene. proach is better at the p=0.1% level (sign test) 7. RELATED WORK irregularity by increasing demand (and hence prices)for these Our research is related to the work in online recommende products n generate same satisfaction levels by intro- systems, in particular, the content-based systems that rec- To test if amend items similar to those that a user liked in the past lucing diversity artificially, we also generated an additional (e.g,(201). Content-based systems learn user preferences from “ diversity ranking” using combined criteria of" pnce”and"ho el class. We interlaced the t otels with“ the lowest the individual-level profiles elicited from users explicitly(e. g through questionnaires or identified transactional behaviors). To price"and the top-5 hotels with "the highest ratings. The compute a"content-based weight vector", a variety of techniques omparison of this "diversity-enabled'" ranking against our algo- were used, such as the Rocchio algorithm, Bayesian classifiers rithm also indicated that surplus-based diversity is better than an artificially-introduced diversity metric: In all cases, users and Winnow algorithm [1]. Our research also leverages work on trongly preferred the surplus-based ranking, as shown in the 2Of course. we substit business traveler'and“ New York”with last column of table 2 335
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 ROMANCE 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 ROMANCE 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 BUSINESS 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 BUSINESS 0 01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 FAMILY 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 FAMILY 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 FRIENDS 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 FRIENDS Figure 3: Sensitivity towards different hotel characteristics for Romance, Business, Family, and Friends-getaway travelers. We present absolute values, without polarity, to illustrate sensitivity on different product characteristics. Figure 4: Sensitivity of different age groups to review rating & review count: Ages groups 18-24 and 25- 34 pay more attention to online reviews compared to other age groups. City Business Romance Family Friends New York 77% 67% 81% 80% Los Angeles 70% 65% 78% 69% D.C. 79% 63% 70% 75% Orlando 87% 85% 91% 84% Table 3: Personalized surplus-based rankings vs. nonpersonalized, surplus-based rankings. The percentage shows the number of users than indicated preference for the personalized ranking compared to the general, non-personalized surplus-based ranking. Each cell of the table corresponds to N = 200 users, for a total of 3200 comparisons. In all cases, the personalized approach is better at the p = 0.1% level (sign test). irregularity by increasing demand (and hence prices) for these products. To test if we can generate same satisfaction levels by introducing diversity artificially, we also generated an additional “diversity ranking” using combined criteria of “price” and “hotel class.” We interlaced the top-5 hotels with “the lowest price” and the top-5 hotels with “the highest ratings.” The comparison of this “diversity-enabled” ranking against our algorithm also indicated that surplus-based diversity is better than an artificially-introduced diversity metric: In all cases, users strongly preferred the surplus-based ranking, as shown in the last column of Table 2. Personalized rankings vs. rankings using average consumer surplus: We generated a few personalized rankings for different cities based on consumer-specific attributes, such as travel purpose. We conducted blind comparisons in a pair-wise fashion, based on 200 anonymous AMT users, for each comparison. Since we did not know the demographics of the users, we asked the MTurk workers to “select the hotel ranking you would prefer to use while trying to help a business traveler to book a hotel in New York” 12. Based on the user responses, customers strongly preferred the personalized ranking that was tailored for a particular travel purpose using our technique (p = 0.01, sign test). For example, in our NYC experiment, 77% customers indicated their preferences towards the business-oriented ranking (ranking tailored for business travelers) rather than the average-level ranking, and 81% customers did so towards the family-oriented ranking (ranking tailored for family trip travelers). The results were similar across all tests. Qualitative analysis: When we asked users’ opinions for these comparisons, users did not bring up the issue of diversity. This was expected, as even our non-personalized rankings were already diverse. Instead, we found customers considering the context and expectations for a given trip. For example, users indicated that hotels for business trips do not necessarily need to be luxury, but need to provide a quiet business environment and easy access to highway or public transportation. On the other hand, for romantic trips, users strongly preferred the recommendations for luxury services with higher class rating. These results highly dovetail with our empirical estimation, suggesting that our ranking model indeed captures consumers’ real purchase motivation behind the scene. 7. RELATED WORK Our research is related to the work in online recommender systems, in particular, the content-based systems that recommend items similar to those that a user liked in the past (e.g., [20]). Content-based systems learn user preferences from the individual-level profiles elicited from users explicitly (e.g., through questionnaires or identified transactional behaviors). To compute a “content-based weight vector”, a variety of techniques were used, such as the Rocchio algorithm, Bayesian classifiers, and Winnow algorithm [1]. Our research also leverages work on 12Of course, we substituted “business traveler’ and “New York” with the appropriate values for each comparison. WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 335
WWw 2011-Session: E-commerce March 28-April 1, 2011, Hyderabad, India nsumer opinions from online reviews [23, 26]. In our References rough user w users 1 ADOMAvIcIUS, G, AND TUZHILIN, A. Toward the generation of to online reviews 2, and demonstrate how to extract of-the-art and possibl aphic-specific preferences. Other studies proposed popularity with user feedback or social 2 ARCHAK, N, GHOSE, A, AND IPEIROTIS, P. G. Show me the refine search results [4, 14 1Dmpm时ymm []BALKE, W.-T, AND GONTZER, U. Multi-objective query processing for 8. DISCUSSION AND FUTURE WORK on Very Large Data Bases (VLDB)(2004), pp. 936-947 We presented a ranking algorithm that uses a behavioral model of consumers. based on utility maximization. The model generates an estimate of how much each product characteristic contributes to the product's overall utility, and estimates the tion. RAND Journal of Economics 25(1994), 242-262 [G] BERRY, S, LEVINSOHN, J, AND PAKES, A Automobile prices in mark sensitivity of consumers to changes in various product character- quilibrium. Econometrica 63(1995), 841-890 istics. The estimation models are privacy-friendly as they do not [7] BERRY, S, AND PAKES, A. The pure characteristics demand model require individual consumer data but rather rely on aggregate nternational Economic Review 48(2007), 1193-1225 data. Based on the generated models, we can estimate the sur- [8]CHEVALIER, J. A, AND GOOLSBEE, A. Me plus that each product generates for each consumer, and build ankings that capture the user preferences. We demonstrated 10m6mmm22102 Examining the through extensive user studies, that our ranking schemes are better than any of the existing baselines, We also showed that personalized surplus-based rankings are even better than the [10 GHOSE, A,, AND IPEIROTIS, P. G. Estimating the helpfulness and reviews: Mining text and reviewer char- non-personalized surplus-based rankings. By doing so, we are acteristics. IEEE TKDE (2010) ble to target at each individual customer, and offer products [11] GHOSE, A, IP with the "best value for money" in response to consumer queries. sing econometrics: A case study on reputation systems. In ACL We should also note that our ranking scheme is "causal. " in the sense that the model can predict what " should" happen [12] HANSEN, L. Large sample properties of generalized method of mo- when we observe changes in the market. For example, when [13 HECKMAN, J. Instrumental variables: A study of implicit behavioral mers'demand for the product. Also, we can dynamically y options us,m四p要mhm we see a new product in the marketplace, we can rank it b al of hu- rving its characteristics, without waiting to see the ement and 15] LANCASTER, K. Consumer Demand: A New Approach. For example, if we observe a price change, or if we observe that a niversity Press. New York. 1971 hotel closes its pool for renovations, we can adjust immediately [16] LI, B. GHOSE, A, AND IPEIROTIS, P. G. Stay elsewhere the surplus values and re-estimate the ranking classification. In WebDB (2008) Also, in order to better understand the antecedents of con- 17 MARSHALL, A. Principles of Economics, Eighth ed. Macmillan and sumers decisions, future work can look not only at transaction data but also into their browsing history and learning behav [18 McFADDEN, D. Conditional Logit Analysis of Qualitative Choic For example, our current model assumes that consumers are gaging into optimal utility maximizing behavior. However 19 MCFADDEN, D, AND TRAIN, K. Mixed MNL models of disc 5(2000),447-470 this is not always true, as some consumers are more thorough [20] MooNEY, R, AND RoY, L. Content-based book recommending using we can build models that explicitly take into consideration the ender Systems: Algorithms and Evaluation (1999 that some users are 21 NELDER, AND MEAD, R. A simplex method for function mini- simply engage into" satisficing. "It would be also interested in ization. The Computer Journal 7, 4(1965) mining the difference in the conversion rate of users. when ge-level search. SIGMOD Record 37, 4(2008), 48-54. presented with surplus-based rankings 23]P ment analysis. Foun- uct search through the "econom ations and Trends in Information Retrieval 2, 1-2(2008) of consumer behavior, we can leverage micro-economic theory 24 ROSEN, S. Hedonic prices and implicit markets: Product differentia- and many theoretical models that have been developed over ion in pure competition. J. of Political Econ. 82, 1(1974) the years, which try to capture the decision-making process and with an application to personal computers. FR OS-09, 2008 of humans. Economic theory provides a very solid basis upon [26] YE, Q 26 YE, Q, LAW, R, AND GU, B. The impact of online which we can build further computer science research, which has a different focus than economic research. Our example 27] YEE, K.-P., SwEARINGEN, K, LI, K, AND HEARST, M. Faceted meta- illustrating: while economists have been building utility models data for image search and browsing. In CHI (2003), pp. 401-408 for their goal was to estimate demand for products and the notion of surplus was just"a to an end"and had of value by itself. By focusing on product ranking, w howed how surplus can improve product search. Our experi- mental results demonstrated a significant improvement in user atisfaction. Other economic models(e. g, measuring the utility of product bundles) can also be directly used in consumer-faci applications on the Web(e. g, search for"product bundles"in- stead of simple products). We are very optimistic that this interdisciplinary research direction can generate very interesting results in the future
learning consumer opinions from online reviews [23, 26]. In our work, through user modeling, we identify how users behave as a response to online reviews [2], and demonstrate how to extract demographic-specific preferences. Other studies proposed to combine popularity with user feedback or social annotations to refine search results [4, 14]. 8. DISCUSSION AND FUTURE WORK We presented a ranking algorithm that uses a behavioral model of consumers, based on utility maximization. The model generates an estimate of how much each product characteristic contributes to the product’s overall utility, and estimates the sensitivity of consumers to changes in various product characteristics. The estimation models are privacy-friendly as they do not require individual consumer data but rather rely on aggregate data. Based on the generated models, we can estimate the surplus that each product generates for each consumer, and build rankings that capture the user preferences. We demonstrated, through extensive user studies, that our ranking schemes are better than any of the existing baselines. We also showed that personalized surplus-based rankings are even better than the non-personalized surplus-based rankings. By doing so, we are able to target at each individual customer, and offer products with the “best value for money” in response to consumer queries. We should also note that our ranking scheme is “causal,” in the sense that the model can predict what “should” happen when we observe changes in the market. For example, when we see a new product in the marketplace, we can rank it by simply observing its characteristics, without waiting to see the consumers’ demand for the product. Also, we can dynamically change the rankings as a reaction to changes in the products. For example, if we observe a price change, or if we observe that a hotel closes its pool for renovations, we can adjust immediately the surplus values and re-estimate the rankings. Also, in order to better understand the antecedents of consumer’s decisions, future work can look not only at transaction data but also into their browsing history and learning behavior. For example, our current model assumes that consumers are engaging into optimal utility maximizing behavior. However, this is not always true, as some consumers are more thorough than others in their search. By leveraging browsing histories, we can build models that explicitly take into consideration the fact that some users are “utility optimizers” and some others simply engage into “satisficing.” It would be also interested in examining the difference in the conversion rate of users, when presented with surplus-based rankings. By examining product search through the “economic lens” of consumer behavior, we can leverage micro-economic theory and many theoretical models that have been developed over the years, which try to capture the decision-making process of humans. Economic theory provides a very solid basis upon which we can build further computer science research, which has a different focus than economic research. Our example is illustrating: while economists have been building utility models for years, their goal was to estimate demand for products and the notion of surplus was just “a means to an end” and never had of value by itself. By focusing on product ranking, we showed how surplus can improve product search. Our experimental results demonstrated a significant improvement in user satisfaction. Other economic models (e.g., measuring the utility of product bundles) can also be directly used in consumer-facing applications on the Web (e.g., search for “product bundles” instead of simple products). We are very optimistic that this interdisciplinary research direction can generate very interesting results in the future. References [1] Adomavicius, G., and Tuzhilin, A. Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE TKDE 17 (2005), 734–749. [2] Archak, N., Ghose, A., and Ipeirotis, P. G. Show me the money!: deriving the pricing power of product features by mining consumer reviews. In KDD (2007), pp. 56–65. [3] Balke, W.-T., and Guntzer, U. ¨ Multi-objective query processing for database systems. In Proceedings of 28th International Conference on Very Large Data Bases (VLDB) (2004), pp. 936–947. [4] Bao, S., Wu, X., Fei, B., Xue, G., Su, Z., and Yu, Y. Optimizing web search using social annotations. In WWW (2007). [5] Berry, S. Estimating discrete choice models of product differentiation. RAND Journal of Economics 25 (1994), 242–262. [6] Berry, S., Levinsohn, J., and Pakes, A. Automobile prices in market equilibrium. Econometrica 63 (1995), 841–890. [7] Berry, S., and Pakes, A. The pure characteristics demand model. International Economic Review 48 (2007), 1193–1225. [8] Chevalier, J. A., and Goolsbee, A. Measuring prices and price competition online: Amazon.com and BarnesandNoble.com. Quantitative Marketing and Economics 1, 2 (2003), 203–222. [9] Forman, C., Ghose, A., and Wiesenfeld, B. Examining the relationship between reviews and sales: the role of reviewer identity disclosure in electronic markets. ISR 19, 3 (2008), 291–313. [10] Ghose, A., , and Ipeirotis, P. G. Estimating the helpfulness and economic impact of product reviews: Mining text and reviewer characteristics. IEEE TKDE (2010). [11] Ghose, A., Ipeirotis, P., and Sundararajan, A. Opinion mining using econometrics: A case study on reputation systems. In ACL (2007). [12] Hansen, L. Large sample properties of generalized method of moments estimators. Econometrica 50, 4 (1982), 1029–1054. [13] Heckman, J. Instrumental variables: A study of implicit behavioral assumptions used in making program evaluations. Journal of Human Resources 32, 3 (1997), 441–462. [14] Jin, R., Valizadegan, H., and Li, H. Ranking refinement and its application to information retrieval. In WWW (2008). [15] Lancaster, K. Consumer Demand: A New Approach. Columbia University Press, New York, 1971. [16] Li, B., Ghose, A., and Ipeirotis, P. G. Stay elsewhere? improving local search for hotels using econometric modeling and image classification. In WebDB (2008). [17] Marshall, A. Principles of Economics, Eighth ed. Macmillan and Co., London, 1926. [18] McFadden, D. Conditional Logit Analysis of Qualitative Choice Behavior. Academic Press, New York, 1974. [19] McFadden, D., and Train, K. Mixed MNL models of discrete response. Journal of Applied Econometrics 15, 5 (2000), 447–470. [20] Mooney, R., and Roy, L. Content-based book recommending using learning for text categorization. In ACM SIGIR Workshop Recommender Systems: Algorithms and Evaluation (1999). [21] Nelder, J. A., and Mead, R. A simplex method for function minimization. The Computer Journal 7, 4 (1965). [22] Nie, Z., Wen, J.-R., and Ma, W.-Y. Webpage understanding: beyond page-level search. SIGMOD Record 37, 4 (2008), 48–54. [23] Pang, B., and Lee, L. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval 2, 1-2 (2008). [24] Rosen, S. Hedonic prices and implicit markets: Product differentiation in pure competition. J. of Political Econ. 82, 1 (1974), 34–55. [25] Song, M. A hybrid discrete choice model of differentiated product demand with an application to personal computers. FR 08-09, 2008. [26] Ye, Q., Law, R., and Gu, B. The impact of online user reviews on hotel room sales. Int. J. of Hosp. Mgmnt. 28, 1 (2009), 180–182. [27] Yee, K.-P., Swearingen, K., Li, K., and Hearst, M. Faceted metadata for image search and browsing. In CHI (2003), pp. 401–408. WWW 2011 – Session: E-commerce March 28–April 1, 2011, Hyderabad, India 336