《电子商务 E-business》阅读文献：Recommender system architecture for adaptive green marketing

Expert Systems with Applications 38(2011)9696-9703 Contents lists available at Science Direct Expert Systems with Applications ELSEVIER journalhomepagewww.elsevier.com/locate/eswa Recommender system architecture for adaptive green marketing Ying-Lien Lee.*, Fei-Hui Huang b Department of Industrial Engineering and Management, Chaoyang University of Technology, Taichung County 413, Taiwan Department of marketing and Distribution Management, Oriental Institute of Technology, Taipei County 220, Taiwan ARTICLE INFO A BSTRACT Green marketing has become an important method for companies to remain profitable and competitive Its are more concerned about environmental issues. however most online shopping environments do not consider product greenness in their recommender systems or other shop- Fuzzy inference system ping tools. This paper aims to propose the use of recommender systems to aid the green shopping proce technique called foot-in-the-door(FITD). In this study, the architecture of a recommender system for green consumer electronics is proposed. Customers' decision making process is modeled with an adaptive fuzzy inference system in which the input variables are the degrees of output variables are the estimated rating data. The architecture has three types of recommendation: information filtering, candidate expansion, and crowd recommendation. Ad hoc customization can be applied to tune the recommendation results. The findings are reported in two parts. The first part describes the potentials of using recommender systems in green marketing and the promotion of green consumerism; the second part describes the proposed recommender system architecture using green lectronics as the context. Discussion of the proposed architecture and comparison with other re also included in this part. The proposed architecture provides a capable platform for person- keting by offering customers shopping advices tailored to their preferences and for the erism e 2011 Elsevier Ltd. All rights reserved. 1 Introduction recommends blogs a rater might be interested in. The domain of recommender systems is not limited to the famous instances men- Recommender systems have become an important technology tioned above Recommender systems for news, web pages, jokes for electronic commerce on many fronts( Bose, 2009: Kauffman& academic articles, consumer electronics, restaurants, and a pleth Walden, 2001). It can filter for online shoppers the vast amount ora of other subject matters, have been researched and imple of information, saving the customers from the information over- mented(Adomavicius Tuzhilin, 2005: Iijima Ho, 2007) ad problem(Chen, Shang, Kao, 2009). It can be a decision aid However, to our knowledge few researches have dealt with rec for customers who are challenged when they are in the market ommender system of green product. for unfamiliar products. It can be a strategic marketing platform Green product is increasingly important in our global village n which online venders can personalize promotions and sales the general public is becoming more concerned of our i for each customer(Chen, 2008: Shih, Chiu, Hsu, lin, 2002). the planet. Driven by this trend, companies have been trying to de- ems have been vigorously researched and sign and manufacture greener products, and have been trying to developed in the fields of academia and business. Some notable promote their products and brand images by communicating their examples include apple Inc's Genius of iTunes that make music greenness to the customers via a variety of channels. Yet, eco- recommendations, University of Minnesotas MovieLens and labeling remains one of the fundamental ways to inform the cus- Netflix's Cinematch recommend movie titles, Amazon. coms tomers how green their products are and in what respect their ecommender system that generates recommendations of an products are green. Eco-labels, usually issued by third-party orga assortment of products, and Outbrain coms blog rating widget that nizations, are textual or graphical presentations of the environ- mental characteristics of a product, which can be found on the product itself, on the packaging, or in the manual. Examples of eco-labels include Green Seal, Energy Star, and WEEE (Waste Elec A*Corresponding author. Address: No. 168. Jifong E Rd, Wufong Township, trical and Electronic Equipment Directive). Studies have shown chung County 413, Taiwan. Tel: +886 4 23323000: fax: +886 4 2374232 E-mailaddressyinglienlee@gmail.com(y.-lLee). that public education campaign is one of the key determinants of 0957-4174 front matter o 2011 Elsevier Ltd. All rights reserved o:10.1016/eswa2011.01.164

Recommender system architecture for adaptive green marketing Ying-Lien Lee a,⇑ , Fei-Hui Huang b aDepartment of Industrial Engineering and Management, Chaoyang University of Technology, Taichung County 413, Taiwan bDepartment of Marketing and Distribution Management, Oriental Institute of Technology, Taipei County 220, Taiwan article info Keywords: Green marketing Green consumerism Recommender system Fuzzy inference system abstract Green marketing has become an important method for companies to remain profitable and competitive as the public and governments are more concerned about environmental issues. However, most online shopping environments do not consider product greenness in their recommender systems or other shop￾ping tools. This paper aims to propose the use of recommender systems to aid the green shopping process and to promote green consumerism basing upon the benefits of recommender systems and a compliance technique called foot-in-the-door (FITD). In this study, the architecture of a recommender system for green consumer electronics is proposed. Customers’ decision making process is modeled with an adaptive fuzzy inference system in which the input variables are the degrees of price, feature, and greenness and output variables are the estimated rating data. The architecture has three types of recommendation: information filtering, candidate expansion, and crowd recommendation. Ad hoc customization can be applied to tune the recommendation results. The findings are reported in two parts. The first part describes the potentials of using recommender systems in green marketing and the promotion of green consumerism; the second part describes the proposed recommender system architecture using green consumer electronics as the context. Discussion of the proposed architecture and comparison with other systems are also included in this part. The proposed architecture provides a capable platform for person￾alized green marketing by offering customers shopping advices tailored to their preferences and for the promotion of green consumerism. 2011 Elsevier Ltd. All rights reserved. 1. Introduction Recommender systems have become an important technology for electronic commerce on many fronts (Bose, 2009; Kauffman & Walden, 2001). It can filter for online shoppers the vast amount of information, saving the customers from the information over￾load problem (Chen, Shang, & Kao, 2009). It can be a decision aid for customers who are challenged when they are in the market for unfamiliar products. It can be a strategic marketing platform on which online venders can personalize promotions and sales for each customer (Chen, 2008; Shih, Chiu, Hsu, & Lin, 2002). Recommender systems have been vigorously researched and developed in the fields of academia and business. Some notable examples include Apple Inc.’s Genius of iTunes that make music recommendations, University of Minnesota’s MovieLens and Netflix’s Cinematch that recommend movie titles, Amazon.com’s recommender system that generates recommendations of an assortment of products, and Outbrain.com’s blog rating widget that recommends blogs a rater might be interested in. The domain of recommender systems is not limited to the famous instances men￾tioned above. Recommender systems for news, web pages, jokes, academic articles, consumer electronics, restaurants, and a pleth￾ora of other subject matters, have been researched and imple￾mented (Adomavicius & Tuzhilin, 2005; Iijima & Ho, 2007). However, to our knowledge, few researches have dealt with rec￾ommender system of green product. Green product is increasingly important in our global village as the general public is becoming more concerned of our impact on the planet. Driven by this trend, companies have been trying to de￾sign and manufacture greener products, and have been trying to promote their products and brand images by communicating their greenness to the customers via a variety of channels. Yet, eco￾labeling remains one of the fundamental ways to inform the cus￾tomers how green their products are and in what respect their products are green. Eco-labels, usually issued by third-party orga￾nizations, are textual or graphical presentations of the environ￾mental characteristics of a product, which can be found on the product itself, on the packaging, or in the manual. Examples of eco-labels include Green Seal, Energy Star, and WEEE (Waste Elec￾trical and Electronic Equipment Directive). Studies have shown that public education campaign is one of the key determinants of 0957-4174/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.01.164 ⇑ Corresponding author. Address: No. 168, Jifong E. Rd., Wufong Township, Taichung County 413, Taiwan. Tel.: +886 4 23323000; fax: +886 4 23742327. E-mail address: yinglienlee@gmail.com (Y.-L. Lee). Expert Systems with Applications 38 (2011) 9696–9703 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

Y-L Lee, F-H. Huang/ Expert Systems with Applications 38(2011)9696-9703 96 uccessful eco-labeling programs(Malcohn, Paulos, Stoeckle 2 Related work Wang,1994) q Public education campaign of eco-labeling programs can be 2.1. Recommender system le via methods such as media coverage, regulation, promotion, school curriculum, and so on. This research proposes using recom- Recommender systems have a variety of forms with different nender systems, in addition to these methods, as a means to edu- functions( Manouselis Costopoulou, 2007: Wan, Menon, &Rama cate and inform on-line customers. The justification of such prasad, 2007 ). Therefore, it warrants a clear definition of the kind proposal relies on two of the primary functions of a recommender of recommender systems this paper is dealing with Schein, Pope- system: information filtering and candidate expansion. When cus- scul, Ungar, and Pennock(2005)define recommender systems as tomers are confronted with a flood of products or with unfamiliar the following:"Recommender systems suggest items of interest products, they may have difficulty in making a shopping decision. to users based on their explicit and implicit preferences, the pref Based upon what they have purchased before, a recommender sys- erences of other users, and user and item attributes". This defini- tem can help the customers by filtering out items that are unlikely tion points out the fundamental parts and necessary input and to be preferred. For example, Amazon. com,s recommender syster output data of a recommender system. First, a recommender sys generates a personalized list of recommended products each time a tem needs data of preferences from single user or multiple users. customer visits their web site. As to candidate expansion, when a The system can explicitly elicit preferences from users by asking mender system can ensure that other good candidates are included ences from past transactions(Resnick Varian, 1997).Second in the consideration set by finding related products based upon the recommender system requires attributes of users and items. product under consideration. Take Amazon coms recommender Manouselis and Costopoulou(2007)refer to these two sets of attri- system for example again. When a customer is looking at the cat- butes as"user model"and"domain model", respectively. Several alog page of a product, the recommender system recommends representations can be used as user models, such as per user prod items similar to the current item. Tapping into the capabilities of uct ratings, demographic attributes, transaction histories, and so information filtering and candidate expansion, recommender sys- on. On the other hand, domain models can be represented as char tems can be transformed into a green product advocate informing acteristics of products and as derived attributes such as taxono- the customers of available choices that are greener. a re mies, hierarchies, and ontologies. Both models may utilize the mender system of green products can also sieve through a set of acquired user preferences to derive their own data. products to retrieve only the items matching an implicitly or The core of a recommender system is the mechanism of sugges- explicitly degree of greenness designated by a customer. Such sys- tion generation based upon the user model and domain model. The tem can also find other products whose greenness and other as- mechanism can be formulated as follows(Adomavicius Tuzhilin ects are comparable based upon a product under consideration. 2005): Let C be the set of all customers and P be the set of all prod- The reduced effort in the decision making process may enhance ucts that a recommender knows of. In addition, let U(c, p) be the he quality and users'satisfaction of the decision(Haubl Trifts, utility function that associates(c, p) pairs with utility values which 2000), which in turn will make green shopping a more enjoyable can be ratings, profits, or some other measurements. The objective experience of a recommender system is to find a set of items pE P such that The adaptability of a recommender system can also contribute U(c p)is maximized for a customer. The mathematical formulation to the promotion of green consumerism by using a techni is as follows: alled foot-in-the-door (FiTD) technique(Freedman Fraser 1966). FITD is a compliance technique in which a person is more cEC, p=argmaxU(c, p) likely to accept a larger request if this request is preceded by smaller request. The technique is also found to be effective in com- In the formulation, arg max"means"the argument of the puter-mediated communication(CMC) in addition to face-to-face maximum telephone communications(Gueguen, 2002 ). In a recommender Recommender systems can be generally classified into three system of green products, items with higher degree of greenness categories according to the mechanism of recommendation gener- and with comparable or equal degrees of price and feature can ation(Adomavicius Tuzhilin, 2005; Schein et al, 2005: (1)Con- be first recommended to a user who is reluctant to buy green prod tent-based systems recommend items that are similar to the ones a ucts. Appropriate feedback should be given to the user about the user preferred in the past. (2)Collaborative systems recommend environmental contribution of the purchase one has made The de- items that other like-minded users preferred in the past. (3)Hybrid gree of greenness of the recommended items in the future can be systems recommend items by combining content-based and col- adjusted accordingly if the users'purchasing transactions reflect laborative methods in recommendation generation. vicius and Tuzhilin(2005) out, content-based The goal of this paper is to develop a recommender system and collaborative systems have some challenges to be dealt with architecture for green consumer electronics. Instead of simply add- For content-based systems, the first problem is"limited content ing an additional green attribute to the conventional recommender analysis", in which case the recommendation is limited by the fea- systems, the architecture uses an adaptive behavioral agent to find tures associated with the items. However, some features are hard- the products of a certain degree of greenness according to users' er to extract than others are. For example, extracting features from behaviors. The agent uses an adaptive fuzzy inference system to textual information is easier than from multimedia data. Also, learn users'behavior over time with a basic assumption that a items that are identical in terms of features are indistinguishable ilateral relationship of either symbiosis or antibiosis exists be- The second problem is overspecialization, in which case the sys- tween the pairs of price vs feature price vs greenness, and feature tem can only recommend items that are similar to items a user s greenness. liked in the past. In other words, the lack of diversity may jeopar- The rest of this paper is organized as follows. The nex dize the practicality of a recommender system. The third problem gives brief review of recommender systems and fuzzy is"new user problem", in which case a user is unable to get reli- tems. The proposed architecture is presented and disci able recommendations until a sufficient amount of transactions Section 3. Conclusions and future research directions are presented are present for the recommender system to learn about the users in the final section

successful eco-labeling programs (Malcohn, Paulos, Stoeckle, & Wang, 1994). Public education campaign of eco-labeling programs can be done via methods such as media coverage, regulation, promotion, school curriculum, and so on. This research proposes using recom￾mender systems, in addition to these methods, as a means to edu￾cate and inform on-line customers. The justification of such proposal relies on two of the primary functions of a recommender system: information filtering and candidate expansion. When cus￾tomers are confronted with a flood of products or with unfamiliar products, they may have difficulty in making a shopping decision. Based upon what they have purchased before, a recommender sys￾tem can help the customers by filtering out items that are unlikely to be preferred. For example, Amazon.com’s recommender system generates a personalized list of recommended products each time a customer visits their web site. As to candidate expansion, when a customer is evaluating the decision to buy a product, a recom￾mender system can ensure that other good candidates are included in the consideration set by finding related products based upon the product under consideration. Take Amazon.com’s recommender system for example again. When a customer is looking at the cat￾alog page of a product, the recommender system recommends items similar to the current item. Tapping into the capabilities of information filtering and candidate expansion, recommender sys￾tems can be transformed into a green product advocate informing the customers of available choices that are greener. A recom￾mender system of green products can also sieve through a set of products to retrieve only the items matching an implicitly or explicitly degree of greenness designated by a customer. Such sys￾tem can also find other products whose greenness and other as￾pects are comparable based upon a product under consideration. The reduced effort in the decision making process may enhance the quality and users’ satisfaction of the decision (Häubl & Trifts, 2000), which in turn will make green shopping a more enjoyable experience. The adaptability of a recommender system can also contribute to the promotion of green consumerism by using a technique called foot-in-the-door (FITD) technique (Freedman & Fraser, 1966). FITD is a compliance technique in which a person is more likely to accept a larger request if this request is preceded by a smaller request. The technique is also found to be effective in com￾puter-mediated communication (CMC) in addition to face-to-face or telephone communications (Guéguen, 2002). In a recommender system of green products, items with higher degree of greenness and with comparable or equal degrees of price and feature can be first recommended to a user who is reluctant to buy green prod￾ucts. Appropriate feedback should be given to the user about the environmental contribution of the purchase one has made. The de￾gree of greenness of the recommended items in the future can be adjusted accordingly if the users’ purchasing transactions reflect acceptance or rejection of the items. The goal of this paper is to develop a recommender system architecture for green consumer electronics. Instead of simply add￾ing an additional green attribute to the conventional recommender systems, the architecture uses an adaptive behavioral agent to find the products of a certain degree of greenness according to users’ behaviors. The agent uses an adaptive fuzzy inference system to learn users’ behavior over time with a basic assumption that a bilateral relationship of either symbiosis or antibiosis exists be￾tween the pairs of price vs. feature, price vs. greenness, and feature vs. greenness. The rest of this paper is organized as follows. The next section gives brief review of recommender systems and fuzzy inference systems. The proposed architecture is presented and discussed in Section 3. Conclusions and future research directions are presented in the final section. 2. Related work 2.1. Recommender system Recommender systems have a variety of forms with different functions (Manouselis & Costopoulou, 2007; Wan, Menon, & Rama￾prasad, 2007). Therefore, it warrants a clear definition of the kind of recommender systems this paper is dealing with. Schein, Pope￾scul, Ungar, and Pennock (2005) define recommender systems as the following: ‘‘Recommender systems suggest items of interest to users based on their explicit and implicit preferences, the pref￾erences of other users, and user and item attributes’’. This defini￾tion points out the fundamental parts and necessary input and output data of a recommender system. First, a recommender sys￾tem needs data of preferences from single user or multiple users. The system can explicitly elicit preferences from users by asking them to rate some items, or implicitly by inferring their prefer￾ences from past transactions (Resnick & Varian, 1997). Second, a recommender system requires attributes of users and items. Manouselis and Costopoulou (2007) refer to these two sets of attri￾butes as ‘‘user model’’ and ‘‘domain model’’, respectively. Several representations can be used as user models, such as per user prod￾uct ratings, demographic attributes, transaction histories, and so on. On the other hand, domain models can be represented as char￾acteristics of products and as derived attributes such as taxono￾mies, hierarchies, and ontologies. Both models may utilize the acquired user preferences to derive their own data. The core of a recommender system is the mechanism of sugges￾tion generation based upon the user model and domain model. The mechanism can be formulated as follows (Adomavicius & Tuzhilin, 2005): Let C be the set of all customers and P be the set of all prod￾ucts that a recommender knows of. In addition, let U(c, p) be the utility function that associates (c, p) pairs with utility values which can be ratings, profits, or some other measurements. The objective of a recommender system is to find a set of items p0 e P such that U(c, p) is maximized for a customer. The mathematical formulation is as follows: 8c 2 C; p0 ¼ arg max p2P Uðc; pÞ In the formulation, ‘‘arg max’’ means ‘‘the argument of the maximum’’. Recommender systems can be generally classified into three categories according to the mechanism of recommendation gener￾ation (Adomavicius & Tuzhilin, 2005; Schein et al., 2005): (1) Con￾tent-based systems recommend items that are similar to the ones a user preferred in the past. (2) Collaborative systems recommend items that other like-minded users preferred in the past. (3) Hybrid systems recommend items by combining content-based and col￾laborative methods in recommendation generation. As Adomavicius and Tuzhilin (2005) point out, content-based and collaborative systems have some challenges to be dealt with. For content-based systems, the first problem is ‘‘limited content analysis’’, in which case the recommendation is limited by the fea￾tures associated with the items. However, some features are hard￾er to extract than others are. For example, extracting features from textual information is easier than from multimedia data. Also, items that are identical in terms of features are indistinguishable. The second problem is overspecialization, in which case the sys￾tem can only recommend items that are similar to items a user liked in the past. In other words, the lack of diversity may jeopar￾dize the practicality of a recommender system. The third problem is ‘‘new user problem’’, in which case a user is unable to get reli￾able recommendations until a sufficient amount of transactions are present for the recommender system to learn about the users’ preferences. Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703 9697

Y-L Lee, F-H Huang/ Expert Systems with Applications 38(2011)9696-9703 Collaborative systems also have the"new user problem"which 3. The architecture is similar to that of content-based systems. Another two problem ollaborative systems are"new item problem"and sparsity. 3. 1. Definitions and assumptions New item problem refers to situations in which the system is un ble to recommend newly added items because these items have The core of the proposed recommender system architecture for fers to situations in which there are few like-minded users, or in learns users'preferences in terms of price, feature, and greenness. which there are items that are preferred by few people Feature is defined as the functional and non-functional characteris tics of a product. Functional characteristics are directly related to 2.2. Fuzzy set theory and fuzzy inference system the utilization of a product. Take digital camera for example, func tional characteristics may include lens magnifying power, sensor A fuzzy set a in x is a set of ordered pairs(Zimmermann, 2001 ): pixel count, battery stamina, and so on. On the other hand,non- functional characteristics are not directly related to the use of a A={(x,(x)kx∈x} oroduct, such as brand, industrial design, emotion invoked by the use of the product, and so on. Greenness is the degree of how green X is a collection of objects. F is called the membership function a product is, and price is the degree of the retail monetary value of a product. The representations of the degrees will be described later. that associates X with the membership space M A fuzzy set whose The adaptive behavioral agent has two assumptions. First, there membership space contains only 0 and 1 is identical to a crisp set In is a bilateral relationship between the pairs of price vs. feature, other words, a fuzzy set is a set without crisp and dichotomous boundary. Consider a set of"tall people". By classical set theory, a price vs greenness, and feature vs. greenness. The relationships berson would have to be taller than a clearly defined height to be are either symbiosis or antibiosis. For example, a product whose feature set is rich and innovative may have a relatively high price ncluded in the tall people set, which is incongruent with our com- However, a product in a competitive market may have an aggres- mon senses. By fuzzy set theory, the person, although slightly short- sively low pricing and still have a rich feature set. Likewise, green- er than a predefined height of tall people, is still considered tall only to a lesser degree. Such reasoning is more compatible with human ess may drive the price higher or lower. For example, green hinking and decision making(Zadeh, 1965 ). roducts are usua result of vigorous research and develop A Fuzzy Inference System(FIS)is a type of Expert System in ment or the use of new material or manufacturing process, which which fuzzy sets and fuzzy rules are used to represent human mean extra costs. Yet, cases, the new design, material, enowledge that is imprecise and uncertain(zimmermann, 2001). tween the three pairs are mutually enhancing or decreasing. So To illustrate the process of an FIS and the concept of fuzzy rules, feature can drive greenness higher or lower, and vice versa. Fig. 1 consider an example from Zimmermann(2001)which is about the setting of a special throttle of a technical process. To achieve illustrates the idea thecorrect"setting of the throttle, the following rules must b The second assumption is that customers make their shopping followed decisions according to the criteria of price, feature and greenness. The inclusion of greenness makes this assumption an unusual one. 1. If Temperature is Low and Pressure is Low, Then set Throttle to However, as the green consumerism is becoming a trend, such Medium assumption is rational. The exclusion of other criteria is a practical 2. If Temperature is Low and Pressure is High, Then set Throttle to simplification of the architecture. Lo 3. 2. Domain model Throttle)are described linguistically such as High or Low instead of The domain model defines the representation of products and as precious values. A user of such FIS would input the current val- the generation of the representations in our recommender architec- the degree of highness and lowness. The FiS will then aggregate stand for the degree of price, feature, and greenness, respectiv p es of Temperature and pressures which are then converted into ture. Each product is represented by three cardinal numbers that the degrees of input variables to infer the appropriate settings of the throttle (i.e. the output). 23. Fuzzy set theory and recommender systems Fuzzy set theory and its extensions have been used used to model users'knowledge of a subject matter to appropriate navigation technique(Kavcic, 2004). Fuzzy se can be used to bridge the gap between linguistic user inputs and precise attributes of products. In a consumer electronics shopping d system(Cao Li, 2007), customers'needs are elicited by answering some questions about their needs and concerns. Based on these answers and some predefined fuzzy sets, the system then computes the weight of each specification to get the suitability Functional scores of products. Fuzzy sets has been applied to model similari- ties between item vs item user vs item and user vs user in movie emendation( Perny Zucker, 2001). Fuzzy sets also has been Fig. 1. The bilateral relationship between the pairs of price vs feature, price vs. used to represent the characteristics of items and users(Wang, greenness, and feature vs greenness. Plus sign means enhancing and minus sign 2004) decreasing

Collaborative systems also have the ‘‘new user problem’’ which is similar to that of content-based systems. Another two problems of collaborative systems are ‘‘new item problem’’ and sparsity. New item problem refers to situations in which the system is un￾able to recommend newly added items because these items have yet to be included in the user preference data. Sparsity problem re￾fers to situations in which there are few like-minded users, or in which there are items that are preferred by few people. 2.2. Fuzzy set theory and fuzzy inference system A fuzzy set e A in X is a set of ordered pairs (Zimmermann, 2001): e A ¼ fðx;le A ðxÞÞjx 2 Xg X is a collection of objects. le AðxÞ is called the membership function that associates X with the membership space M. A fuzzy set whose membership space contains only 0 and 1 is identical to a crisp set. In other words, a fuzzy set is a set without crisp and dichotomous boundary. Consider a set of ‘‘tall people’’. By classical set theory, a person would have to be taller than a clearly defined height to be included in the tall people set, which is incongruent with our com￾mon senses. By fuzzy set theory, the person, although slightly short￾er than a predefined height of tall people, is still considered tall only to a lesser degree. Such reasoning is more compatible with human thinking and decision making (Zadeh, 1965). A Fuzzy Inference System (FIS) is a type of Expert System in which fuzzy sets and fuzzy rules are used to represent human knowledge that is imprecise and uncertain (Zimmermann, 2001). To illustrate the process of an FIS and the concept of fuzzy rules, consider an example from Zimmermann (2001) which is about the setting of a special throttle of a technical process. To achieve the ‘‘correct’’ setting of the throttle, the following rules must be followed: 1. If Temperature is Low and Pressure is Low, Then set Throttle to Medium. 2. If Temperature is Low and Pressure is High, Then set Throttle to Low. The input and output variables (i.e. Temperature, Pressure, and Throttle) are described linguistically such as High or Low instead of as precious values. A user of such FIS would input the current val￾ues of Temperature and Pressures which are then converted into the degree of highness and lowness. The FIS will then aggregate the degrees of input variables to infer the appropriate settings of the throttle (i.e. the output). 2.3. Fuzzy set theory and recommender systems Fuzzy set theory and its extensions have been used to model various aspects of recommender systems. For example, it has been used to model users’ knowledge of a subject matter to select an appropriate navigation technique (Kavcic, 2004). Fuzzy set theory can be used to bridge the gap between linguistic user inputs and precise attributes of products. In a consumer electronics shopping aid system (Cao & Li, 2007), customers’ needs are elicited by answering some questions about their needs and concerns. Based on these answers and some predefined fuzzy sets, the system then computes the weight of each specification to get the suitability scores of products. Fuzzy sets has been applied to model similari￾ties between item vs. item, user vs. item, and user vs. user in movie recommendation (Perny & Zucker, 2001). Fuzzy sets also has been used to represent the characteristics of items and users (Wang, 2004). 3. The architecture 3.1. Definitions and assumptions The core of the proposed recommender system architecture for green consumer electronics is an adaptive behavioral agent that learns users’ preferences in terms of price, feature, and greenness. Feature is defined as the functional and non-functional characteris￾tics of a product. Functional characteristics are directly related to the utilization of a product. Take digital camera for example, func￾tional characteristics may include lens magnifying power, sensor pixel count, battery stamina, and so on. On the other hand, non￾functional characteristics are not directly related to the use of a product, such as brand, industrial design, emotion invoked by the use of the product, and so on. Greenness is the degree of how green a product is, and price is the degree of the retail monetary value of a product. The representations of the degrees will be described later. The adaptive behavioral agent has two assumptions. First, there is a bilateral relationship between the pairs of price vs. feature, price vs. greenness, and feature vs. greenness. The relationships are either symbiosis or antibiosis. For example, a product whose feature set is rich and innovative may have a relatively high price. However, a product in a competitive market may have an aggres￾sively low pricing and still have a rich feature set. Likewise, green￾ness may drive the price higher or lower. For example, green products are usually the result of vigorous research and develop￾ment or the use of new material or manufacturing process, which mean extra costs. Yet, in some cases, the new design, material, or manufacturing process actually saves money. Relationships be￾tween the three pairs are mutually enhancing or decreasing. So, feature can drive greenness higher or lower, and vice versa. Fig. 1 illustrates the idea. The second assumption is that customers make their shopping decisions according to the criteria of price, feature, and greenness. The inclusion of greenness makes this assumption an unusual one. However, as the green consumerism is becoming a trend, such assumption is rational. The exclusion of other criteria is a practical simplification of the architecture. 3.2. Domain model The domain model defines the representation of products and the generation of the representations in our recommender architec￾ture. Each product is represented by three cardinal numbers that stand for the degree of price, feature, and greenness, respectively. Functional Non-functional Fig. 1. The bilateral relationship between the pairs of price vs. feature, price vs. greenness, and feature vs. greenness. Plus sign means enhancing and minus sign decreasing. 9698 Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703

Y-L Lee, F-H Huang/ Expert Systems with Applications 38 (2011)9696-9703 96 699 The degree of price is a representation of how expensive a pro ducers of consumer electronics reveal the greenness of their prod- uct is relative to other products under the same category and the ucts via eco-labels. However, most eco-labels are concerned with Z-score normalization of the products' prices is used to obtain this only some aspects of the overall greenness. For example, Energy degree. Let Pi be the set of all products under category i For a prod- Star reveals the information of energy consumption, and roHS uct pE P the degree of price DP of p is defined as the following hort for Restriction of Hazardous Substances Directive) the infor mation of raw material selection. The proposed architecture uses Analytic Hierarchy Process, or AHP(Saaty, 1987), to elicit experts knowledge of the eco-labels' weights relative to the overall green- Me and Sp, are the means and the standard deviations of the prices ness of a products, as Fig. 2 illustrates. After the process of AHP of products under category i, and Xp is the price of a particular prod calculation, the weights of eco-labels are obtained. Let E=e, ct p. Products with higher DP values are more expensive than ez,.. en be the set of all eco-labels the proposed architecture con those with lower Dp values siders, the vector W is the weights of (e1 e2, .. en). For a product In the same way the degree of feature represents the powerful- PEP, the Eco-label-based degree of greenness DGEco-label of p is ness of a product relatively to other products under the same cat- defined as the following egory. The average of the z-scores of the quantifiable specification of a product category is used to represent this degree. Let S be theDGEco-label set of all quantifiable specifications under category i. For a product PEP, the degree of feature DF of p is defined as the average of the z- scores of s'∈S,i,e, MECO-label and SEco-label are the means and the standard deviations of the eco-label-based of greenness of products under category DF=avg i. XEco-label is the eco ased degree of greenness of a particular product p, which as the following: Myp and Ssp are the means and the standard deviations of the spec- ification s'of products under category i, and X is the specifications XEco-labtlLxw of a particular product p Products witn nigner Dr values are more L is a vector whose n elements are either ero indicatin ifications are more favorable when their numbers are smaller: such whether product p has or does not have ular eco-label specifications need to be converted first. Therefore products whose DGEco-label are also Eco-label Xs=MAXs-Xs where MAXs is the maximal value of specification s 33. User model The degree of greenness is a representation of how product is in relation to other products under the same While DP and dF are readily derivable from the product ion, the degree of greenness is not. Most, if not all, produc t informa- While the domain model described above is a set of attributes the user model is represented by an FIS-based adaptive behavioral agent whose input variables are the attributes of the domain model logs do not have integrated ratings of the greenness of products. and the output variable is the estimated rating of a product. As de- In addition, the assessment of the greenness of a product, scribed in Section 2, an FIS contains fuzzy if-then rules that are de- usually involves a thorough Life Cycle Assessment(LCA), is consuming( Cooper& Fava, 2006). Therefore, the proposed determine the membership functions of an FIS is by observing the tecture uses two approximation methods to represent the degree input and output variables and choosing the membership func- enne tions that fit the variables. The resultant system is a static FIs The first representation is based upon the criteria of Electronic whose membership functions and parameters of the functions Products Environmental Assessment Tool(EPEAT) which is a rating are predetermined. Another method of constructing the fuzzy if- system that gauges the environmental performance of a product then rules is through a soft computing method called Adaptive according to the fifty-one criteria defined in IEEE 1680 standard. Neuro-Fuzzy Inference Systems(ANFIS), in which the rules are ob- EPEAT classifies products satisfying all the twenty-three required tained by training the system with the input and output variables criteria as Bronze Medal. all the twenty-three required criteria (ang. 1993). and 50% of the twenty-eight optional criteria as Silver Medal, and The proposed architecture uses ANFIS and the rating data of all the twenty-three required criteria and 75% of the twenty-eight products to obtain a Fis that represents a users'preference in terms optional criteria as Gold Medal. The proposed architecture uses the of price, feature, and greenness. The architecture trains the anFIs number of satisfied criteria as one of the representations of the de- by feeding it with training data set (p, DP, DF, DG, r IpE P), where gree of greenness. For a product pE P. the EPEAT-based degree of p is a product and r is the rating of a product supplied by the user greenness DGEPEAT of p is defined as the following: The resultant FIS of Pi for a user takes the input of DP, DF, DG) and MPEAT and sEPaT are the means and the standard deviations of the of the gre number of satisfied criteria of products under category i, and Xp is the number of satisfied criteria of a particular product p. The z-score DGEPEAT of a product uses only the data from other prod ucts whose dgEpEar are also EPEAt-based Although EPEAT is an increasingly popular overall greenness Eco-label 2 Eco-label n rating system, its database currently contains only 1175 products. To compensate this limitation, the proposed architecture uses a Fig. 2. The AHP used in the architecture to elicit experts knowledge of the weights complementary degree of greenness based on eco-labels. Most pro- of the eco-labels

The degree of price is a representation of how expensive a prod￾uct is relative to other products under the same category, and the z-score normalization of the products’ prices is used to obtain this degree. Let Pi be the set of all products under category i. For a prod￾uct p e Pi, the degree of price DP of p is defined as the following: DP ¼ Xp MPi SPi MPi and SPi are the means and the standard deviations of the prices of products under category i, and Xp is the price of a particular prod￾uct p. Products with higher DP values are more expensive than those with lower DP values. In the same way, the degree of feature represents the powerful￾ness of a product relatively to other products under the same cat￾egory. The average of the z-scores of the quantifiable specifications of a product category is used to represent this degree. Let S be the set of all quantifiable specifications under category i. For a product p e Pi, the degree of feature DF of p is defined as the average of the z-scores of s0 e S, i.e., DF ¼ avg s02S Xs0 Ms0Pi Ss0Pi Ms0Pi and Ss0Pi are the means and the standard deviations of the spec￾ification s0 of products under category i, and Xs0 is the specification s0 of a particular product p. Products with higher DF values are more powerful than those with lower DF values. Some quantifiable spec￾ifications are more favorable when their numbers are smaller; such specifications need to be converted first. Therefore, Xs0 ¼ MAXs0 Xs0 where MAXs0 is the maximal value of specification s0 . The degree of greenness is a representation of how green a product is in relation to other products under the same category. While DP and DF are readily derivable from the product informa￾tion, the degree of greenness is not. Most, if not all, product cata￾logs do not have integrated ratings of the greenness of products. In addition, the assessment of the greenness of a product, which usually involves a thorough Life Cycle Assessment (LCA), is time￾consuming (Cooper & Fava, 2006). Therefore, the proposed archi￾tecture uses two approximation methods to represent the degree of greenness. The first representation is based upon the criteria of Electronic Products Environmental Assessment Tool (EPEAT) which is a rating system that gauges the environmental performance of a product according to the fifty-one criteria defined in IEEE 1680 standard. EPEAT classifies products satisfying all the twenty-three required criteria as Bronze Medal, all the twenty-three required criteria and 50% of the twenty-eight optional criteria as Silver Medal, and all the twenty-three required criteria and 75% of the twenty-eight optional criteria as Gold Medal. The proposed architecture uses the number of satisfied criteria as one of the representations of the de￾gree of greenness. For a product p e Pi, the EPEAT-based degree of greenness DGEPEAT of p is defined as the following: DGEPEAT ¼ XEPEAT p MEPEAT Pi SEPEAT Pi MEPEAT Pi and SEPEAT Pi are the means and the standard deviations of the number of satisfied criteria of products under category i, and XEPEAT P is the number of satisfied criteria of a particular product p. The z-score DGEPEAT of a product uses only the data from other prod￾ucts whose DGEPEAT are also EPEAT-based. Although EPEAT is an increasingly popular overall greenness rating system, its database currently contains only 1175 products. To compensate this limitation, the proposed architecture uses a complementary degree of greenness based on eco-labels. Most pro￾ducers of consumer electronics reveal the greenness of their prod￾ucts via eco-labels. However, most eco-labels are concerned with only some aspects of the overall greenness. For example, Energy Star reveals the information of energy consumption, and RoHS (short for Restriction of Hazardous Substances Directive) the infor￾mation of raw material selection. The proposed architecture uses Analytic Hierarchy Process, or AHP (Saaty, 1987), to elicit experts’ knowledge of the eco-labels’ weights relative to the overall green￾ness of a products, as Fig. 2 illustrates. After the process of AHP calculation, the weights of eco-labels are obtained. Let E = {e1, e2, ..., en} be the set of all eco-labels the proposed architecture con￾siders, the vector W is the weights of {e1, e2, ..., en}. For a product p e Pi, the Eco-label-based degree of greenness DGEco-label of p is defined as the following: DGEcolabel ¼ XEcolabel p MEcolabel Pi SEcolabel Pi MEcolabel Pi and S Ecolabel Pi are the means and the standard deviations of the Eco-label-based degree of greenness of products under category i. XEcolabel P is the Eco-label-based degree of greenness of a particular product p, which is defined as the following: XEcolabel p ¼ Lp W Lp is a vector whose n elements are either one or zero indicating whether product p has or does not have a particular eco-label. The z-score DGEco-label of a product uses only the data from other products whose DGEco-label are also Eco-label-based. 3.3. User model While the domain model described above is a set of attributes, the user model is represented by an FIS-based adaptive behavioral agent whose input variables are the attributes of the domain model and the output variable is the estimated rating of a product. As de￾scribed in Section 2, an FIS contains fuzzy if-then rules that are de￾rived from human experts’ knowledge of a system. One method to determine the membership functions of an FIS is by observing the input and output variables and choosing the membership func￾tions that fit the variables. The resultant system is a static FIS whose membership functions and parameters of the functions are predetermined. Another method of constructing the fuzzy if￾then rules is through a soft computing method called Adaptive Neuro-Fuzzy Inference Systems (ANFIS), in which the rules are ob￾tained by training the system with the input and output variables (Jang, 1993). The proposed architecture uses ANFIS and the rating data of products to obtain a FIS that represents a users’ preference in terms of price, feature, and greenness. The architecture trains the ANFIS by feeding it with training data set {p, DP, DF, DG, r |p e Pi}, where p is a product and r is the rating of a product supplied by the user. The resultant FIS of Pi for a user takes the input of {DP, DF, DG} and Which eco-label is the most determinant factor of the greenness of a product? Eco-label 1 Eco-label 2 …... Eco-label n Fig. 2. The AHP used in the architecture to elicit experts’ knowledge of the weights of the eco-labels. Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703 9699

Y-L Lee, F-H Huang/ Expert Systems with Applications 38(2011)9696-9703 outputs r, the estimation of r, of product p. The process is illus-. 4. Recommendation generation trated in Fig 3. The use of the obtained Fis will be described later. To minimize the amount of user interaction the architecture Recommendation generation is the process that takes account uses Asynchronous JavaScript and XML (AJAX) to elicit the rating of the input variables(i.e. DP, DF, and dG)and produces an estima data via one click from the user. In addition, the architecture uses tion of the output variables (i.e. r). The proposed architecture ha the ANFis module of the Fuzzy Logic Toolbox of MATLAB as the three types of recommendation: information filtering, candidate core of the adaptive behavioral agent representing users' prefer- expansion, and crowd recommendation. ence. The ANFIS used in the proposed architecture is a five-layer The aim of information filtering is to prevent a customer from network with three input variables (i.e. DP, DF, and DG)and one being overwhelmed by the amount of information, such as unfa output variable(estimated r, or r). Fig 4 illustrates the resultant miliar products or a long list of products to choose from. In the pro- FIS of a sample pro-environment and price-sensitive customer. posed architecture, information filtering is very straightforward in Fig 4a is the rating matrix in which rows are (DP, DF, DG, r. Given that the customers' FIs acts like a filter. When a customer c is the characteristics of this customer, the products with low price browsing products in P. the DP, DF, and dg of each product p in and high degree of greenness are given the highest rating(5 Pi is passed to the Fis of the customer FISc and the one with max stars), while the others are given the lowest rating(1 star). The imal r is recommended to the customer, i.e. surface plot of the estimated output and the inputs is shown in VcEC, P=arg max FIS(p, DPp, DFp, DGp) The recommendation can generate one item(maximal r)or a num- ber of items(e.g. top five rs). The process of recommendation gen- eration is illustrated in Fig. 5 Rating data Product database Ad hoc customization Rule p, DP DF DG rI P∈P catena conversion ANFIS FIS of Best or Top-N database/DP, DF, DG) L customer c User model Fig 3. The generation of user model in the pro Fig. 5. The process of the first type of recommendation: information filtering 113311 131313 33 (b) Fig 4.(a) Sample rating data of a pro-environmental and price-sensitive customer; (b)The surface plot of the estimated output and the inputs

outputs r0 , the estimation of r, of product p. The process is illus￾trated in Fig. 3. The use of the obtained FIS will be described later. To minimize the amount of user interaction, the architecture uses Asynchronous JavaScript and XML (AJAX) to elicit the rating data via one click from the user. In addition, the architecture uses the ANFIS module of the Fuzzy Logic Toolbox of MATLAB as the core of the adaptive behavioral agent representing users’ prefer￾ence. The ANFIS used in the proposed architecture is a five-layer network with three input variables (i.e. DP, DF, and DG) and one output variable (estimated r, or r0 ). Fig. 4 illustrates the resultant FIS of a sample pro-environment and price-sensitive customer. Fig. 4a is the rating matrix in which rows are {DP, DF, DG, r}. Given the characteristics of this customer, the products with low price and high degree of greenness are given the highest rating (5 stars), while the others are given the lowest rating (1 star). The surface plot of the estimated output and the inputs is shown in Fig. 4b. 3.4. Recommendation generation Recommendation generation is the process that takes account of the input variables (i.e. DP, DF, and DG) and produces an estima￾tion of the output variables (i.e. r0 ). The proposed architecture has three types of recommendation: information filtering, candidate expansion, and crowd recommendation. The aim of information filtering is to prevent a customer from being overwhelmed by the amount of information, such as unfa￾miliar products or a long list of products to choose from. In the pro￾posed architecture, information filtering is very straightforward in that the customers’ FIS acts like a filter. When a customer c is browsing products in Pi, the DP, DF, and DG of each product p in Pi is passed to the FIS of the customer FISc, and the one with max￾imal r0 is recommended to the customer, i.e. 8c 2 C; p0 ¼ arg max p2Pi FIScðp;DPp;DFp;DGpÞ The recommendation can generate one item (maximal r0 ) or a num￾ber of items (e.g. top five r0 s). The process of recommendation gen￾eration is illustrated in Fig. 5. {p, DP, DF, DG, r | p ∈ Pi } Fig. 3. The generation of user model in the proposed architecture. Fig. 4. (a) Sample rating data of a pro-environmental and price-sensitive customer; (b) The surface plot of the estimated output and the inputs. FIS of customer c User Model Product database Pi Best or Top-N recommendation {DP, DF, DG} ' r Rule adjustment Criteria designation Criteria conversion Ad hoc customization Non-model criteria Fig. 5. The process of the first type of recommendation: information filtering. 9700 Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703

Although the user model inferred from the rating data is used to ries the product database for products whose fuzzified DP, DF, and represent the users' preference, there are situations when certain DG are the same as those of pn. The process of recommendation eeds cannot be reflected in the user model, for example, when a based upon a reference product is illustrated in Fig. 6. user wants to use the search facility of an online shop to find prod The third type of recommendation, crowd recommendation, ucts of a certain brand or price range. Such situations are handled makes use of the transactions of other customers to generate rec- in the ad hoc customization( the upper block of Fig. 5)in the archi- ommendations. The architecture first clusters the user into differ ecture. There are three types of ad hoc customizations. The first ent groups according to the Fis rules of each customer so that type is criteria designation in which any of the three criteria con- customers in the same group have similar rules. Next, the architec- sidered in the user model is designated by the user in the form ture queries the transaction database to find out what products a of search criteria. The designated criterion or criteria are then con- group of like-minded customers have bought. The recommenda- verted to non-fuzzy rule or rules to replace the corresponding fuz- tion is then presented to a customer in that group based upon zy ones in the user model. This type of customization is suitable for the result of the query. The process of recommendation based upon sers who are looking for products of specific specifications. The intra-crowd similarity is illustrated in Fig. 7. econd type is rule adjustment in which any of the rules in the user When a products' rating data is not enough or when a new cus- model can be adjusted. For example, one may be searching for tomer has just begun to use the system, a quasi-conventional wi en product (mid level greenness) with reasonable price tag dom FIs is used. The rationale is as follows: When price(DP)is low (low price). Such linguistic values are then defuzzified and used, and feature(DFor greenness(DG)is high, it is a great bargain in tandem with the non-adjusted FIS rule in the user model, to fil-(hence favorable). When you pay more for more or less for less. ter the products. The third type is non-model criteria in which cri- it is a neutral condition(DP is high and dF or DG is high; DP is teria not included in the user model are used to filter the products. low and dF or DG is low ) When you pay more for less, it is unde- The non-model criteria are used as-is in the filtering of the recom sirable( dp is high and dF or dG is low ) It is favorable when feature mendation result based on the user model nd greenness are mutually enhancing(DF is high and DG is high) he second type of recommendation, candidate expansion, undesirable when feature is poor and greenness is low (DF is poor serves a different function from the first type. while the first type and dg is low ) It is neutral when the relationship between feature reduces the amount of information, the second type increases the and greenness is antibiosis(DF is poor and dG is high; DF is rich mount of information. The process of candidate expansion is de- and dG is low). Table 1 summarizes these fuzzy rules; Fig. 8 is scribed as the following: When a customer is currently investigat- the implementation of the quasi-conventional wisdom FIS In the ing a product that is satisfactory in terms of the criteria considered, figure, a product with (DP= 2.17, DF=3.85, DG=3.89 is passed to the system can recommend other products that are similar in the FIS and the estimated rating(the output )is 3. 24. terms of the criteria. Let pn be the product that is currently viewed by a customer. The DP, DF, and dG of pn first go through fuzzification 3.5. Discussion that transforms the values into membership degrees and linguistic values. For example, (DP=-2.3, DF=-2.2, DG=-22) is trans- Unlike other recommendation mechanisms, the proposed archi formed into[cheap, feature poor, not green]. Then, the system que- tecture considers relatively few attributes: DP is a normalized The quasi-conventional wisdom FIS. database IF DP IS High AND DF IS Rich THEN RATING IS Neutral (DP, DF, DG) F DP IS High AND DF IS Poor THEN RATING IS Ur P/DP, DF, DG/ F DP IS LOW AND DF IS Rich F DP IS Low AND DF IS Poor THEN RATING IS Neutral F DP IS High AND DG is High THEN RATING IS Neutral values of P linguistic values of p IF DP IS Low AND DG is High THEN RATING IS Favorable Query F DF IS Rich AND DG IS Hig THEN RATING IS Favorable candidates IF DF IS Poor AND DG IS Low THEN RATING IS Undesirable IF DF IS Rich AND DG IS Low THEN RATING IS Neutral Fig. 6. The process of the second type of recommendation: candidate expansion. DF IS Poor AND DG IS High THEN RATING IS Neutral User Clustering User mo Fig. 7. The process of the third type of recommendation: crowd recommendation

Although the user model inferred from the rating data is used to represent the users’ preference, there are situations when certain needs cannot be reflected in the user model, for example, when a user wants to use the search facility of an online shop to find prod￾ucts of a certain brand or price range. Such situations are handled in the ad hoc customization (the upper block of Fig. 5) in the archi￾tecture. There are three types of ad hoc customizations. The first type is criteria designation in which any of the three criteria con￾sidered in the user model is designated by the user in the form of search criteria. The designated criterion or criteria are then con￾verted to non-fuzzy rule or rules to replace the corresponding fuz￾zy ones in the user model. This type of customization is suitable for users who are looking for products of specific specifications. The second type is rule adjustment in which any of the rules in the user model can be adjusted. For example, one may be searching for green product (mid level greenness) with reasonable price tag (low price). Such linguistic values are then defuzzified and used, in tandem with the non-adjusted FIS rule in the user model, to fil￾ter the products. The third type is non-model criteria in which cri￾teria not included in the user model are used to filter the products. The non-model criteria are used as-is in the filtering of the recom￾mendation result based on the user model. The second type of recommendation, candidate expansion, serves a different function from the first type. While the first type reduces the amount of information, the second type increases the amount of information. The process of candidate expansion is de￾scribed as the following: When a customer is currently investigat￾ing a product that is satisfactory in terms of the criteria considered, the system can recommend other products that are similar in terms of the criteria. Let pn be the product that is currently viewed by a customer. The DP, DF, and DG of pn first go through fuzzification that transforms the values into membership degrees and linguistic values. For example, {DP = 2.3, DF = 2.2, DG = 2.2} is trans￾formed into {cheap, feature poor, not green}. Then, the system que￾ries the product database for products whose fuzzified DP, DF, and DG are the same as those of pn. The process of recommendation based upon a reference product is illustrated in Fig. 6. The third type of recommendation, crowd recommendation, makes use of the transactions of other customers to generate rec￾ommendations. The architecture first clusters the user into differ￾ent groups according to the FIS rules of each customer so that customers in the same group have similar rules. Next, the architec￾ture queries the transaction database to find out what products a group of like-minded customers have bought. The recommenda￾tion is then presented to a customer in that group based upon the result of the query. The process of recommendation based upon intra-crowd similarity is illustrated in Fig. 7. When a products’ rating data is not enough or when a new cus￾tomer has just begun to use the system, a quasi-conventional wis￾dom FIS is used. The rationale is as follows: When price (DP) is low and feature (DF) or greenness (DG) is high, it is a great bargain (hence favorable). When you pay more for more or less for less, it is a neutral condition (DP is high and DF or DG is high; DP is low and DF or DG is low). When you pay more for less, it is unde￾sirable (DP is high and DF or DG is low). It is favorable when feature and greenness are mutually enhancing (DF is high and DG is high), undesirable when feature is poor and greenness is low (DF is poor and DG is low). It is neutral when the relationship between feature and greenness is antibiosis (DF is poor and DG is high; DF is rich and DG is low). Table 1 summarizes these fuzzy rules; Fig. 8 is the implementation of the quasi-conventional wisdom FIS. In the figure, a product with {DP = 2.17, DF = 3.85, DG=3.89} is passed to the FIS and the estimated rating (the output) is 3.24. 3.5. Discussion Unlike other recommendation mechanisms, the proposed archi￾tecture considers relatively few attributes: DP is a normalized Pi Product database {DP, DF, DG} pn linguistic values of pn Query Engine Comparable candidates Reference product Fuzzification Fuzzification {DP, DF, DG} linguistic values of Pi Fig. 6. The process of the second type of recommendation: candidate expansion. User Model User Clustering Cluster 1 Cluster 2 Cluster n Transaction Database Recommendation for Cluster 1 Recommendation for Cluster 2 Recommendation for Cluster n Fig. 7. The process of the third type of recommendation: crowd recommendation. Table 1 The quasi-conventional wisdom FIS. IF DP IS High AND DF IS Rich THEN RATING IS Neutral IF DP IS High AND DF IS Poor THEN RATING IS Undesirable IF DP IS Low AND DF IS Rich THEN RATING IS Favorable IF DP IS Low AND DF IS Poor THEN RATING IS Neutral IF DP IS High AND DG IS High THEN RATING IS Neutral IF DP IS High AND DG IS Low THEN RATING IS Undesirable IF DP IS Low AND DG IS High THEN RATING IS Favorable IF DP IS Low AND DG IS Low THEN RATING IS Neutral IF DF IS Rich AND DG IS High THEN RATING IS Favorable IF DF IS Poor AND DG IS Low THEN RATING IS Undesirable IF DF IS Rich AND DG IS Low THEN RATING IS Neutral IF DF IS Poor AND DG IS High THEN RATING IS Neutral Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703 9701

Y-L Lee, F-H Huang/ Expert Systems with Applications 38 (2011)9696-9703 DF385 Fig 8. The rules of the quasi-conventional wisdom FIS. score of price: DG and DF are derived from multiple attributes. The been bought by other customers in the same group. These queries first rationale of the simplified representation is due to the incon- need only one-time scan of the records of the database, which is clusiveness of researches about green consumer behavior. Strau- efficient on modern database management systems. ghan and Roberts(1999)compared the results of past researches The database of the customers' FIS rules has value for marketing udying the relationship between the characteristics of consumers and product design. Given the advances in science and technolog and ecologically conscious consumer behavior(ECCB). They sur- products could have been designed with outstanding features al veyed research that consider the factors of demographic character- still remain fairly green. However, the price tag would have held istics (such as age, gender, income, and education) and customers back. Consumer electronics companies can gain from psychographic characteristics(such as political orientation, altru- the FIs rules the insight of how much more customers are willing m, perceived consumer effectiveness, and environmental con- to pay for extra features and high greenness. cern), and found that most of the results were inconclusive There is little research of recommender system for green prod- equivocal, or mixed. The second rationale of the simplified repre- ucts. Li ao(2008)proposed a green product recommendation sys sentation is to reduce the amount of user intervention. The only tem in which the product database is queried against user user input the proposed architecture is asking for is product rating. designated weights of green criteria. EPEAT is used to represent Yet, the architecture may even risk losing the customers' patience, the greenness of products as well. Liao's system is effective in re not to mention asking customers to fill out questionnaires to gath- ommending green products: however, users are required to input the demographic or psychographic profiles. In addition, the sys- the weights of certain green indices In our system, on the other m would have to ask the customers to fill out the questionnaires hand, users' criteria of the three decision variables are implicitly again every once in a while to see if their profiles change, which is gathered via users'rating of products For users who do not want impractical. The third rationale is privacy issues. Asking customers the hassle, our system can use the quasi-conventional wisdom to reveal so much information might deter them from shopping at FIs to recommend product. Reducing the level of user intervention is important to the success of a recommender system. The pro- The use of normalization score is justified by the ability of the posed architecture is in line with other commercially available rec- user model to work across product category. Normalized scores al- ommender systems, such as Cinematch of Nextflix or Genius of low comparison among products of different categories Without iTunes, because in these systems, rating data is the minimal re- such measures, the recommender system would have to have a quired inpu user model for each product category To enhance the responsiveness of the recommender system, most calculations and fuzzification can be done off ation and fuzzification of the domain model (i.e. (DP, DF, DG) of products) can be done only when the product catalog is changed In this paper, we proposed using recom also be done offline. The only part that must be done online(or based upon thes g process and to promote green consumerism The training of the ANFIS and the clustering of customers'FIS can the real-time) is the recommendation generation. In the information compliance technique. a proposed recommender system architec filtering recommendation, products of a category pass through a ture in the context of green consumer electronics was then de- customers' FIs to generate recommendation; this process is fast ribed and discussed. Previous research on this type of thanks to the efficiency of fuzzy operations. In the candidate recommender system used weights of certain green indices as xpansion recommendation, products of a category for certain fuzzified words. While in the crowd recom the transaction database is queried for other products emendation. puts The architecture proposed in this paper accepts implicit and explicit criteria by modeling user preference and by considering ad hoc modification. The domain model is defined by the normalized

score of price; DG and DF are derived from multiple attributes. The first rationale of the simplified representation is due to the incon￾clusiveness of researches about green consumer behavior. Strau￾ghan and Roberts (1999) compared the results of past researches studying the relationship between the characteristics of consumers and ecologically conscious consumer behavior (ECCB). They sur￾veyed research that consider the factors of demographic character￾istics (such as age, gender, income, and education) and psychographic characteristics (such as political orientation, altru￾ism, perceived consumer effectiveness, and environmental con￾cern), and found that most of the results were inconclusive, equivocal, or mixed. The second rationale of the simplified repre￾sentation is to reduce the amount of user intervention. The only user input the proposed architecture is asking for is product rating. Yet, the architecture may even risk losing the customers’ patience, not to mention asking customers to fill out questionnaires to gath￾er the demographic or psychographic profiles. In addition, the sys￾tem would have to ask the customers to fill out the questionnaires again every once in a while to see if their profiles change, which is impractical. The third rationale is privacy issues. Asking customers to reveal so much information might deter them from shopping at a store. The use of normalization score is justified by the ability of the user model to work across product category. Normalized scores al￾low comparison among products of different categories. Without such measures, the recommender system would have to have a user model for each product category. To enhance the responsiveness of the recommender system, most calculations and fuzzifications can be done offline. The gener￾ation and fuzzification of the domain model (i.e. {DP, DF, DG} of products) can be done only when the product catalog is changed. The training of the ANFIS and the clustering of customers’ FIS can also be done offline. The only part that must be done online (or real-time) is the recommendation generation. In the information filtering recommendation, products of a category pass through a customers’ FIS to generate recommendation; this process is fast thanks to the efficiency of fuzzy operations. In the candidate expansion recommendation, products of a category are queried for certain fuzzified words. While in the crowd recommendation, the transaction database is queried for other products that have been bought by other customers in the same group. These queries need only one-time scan of the records of the database, which is efficient on modern database management systems. The database of the customers’ FIS rules has value for marketing and product design. Given the advances in science and technology, products could have been designed with outstanding features and still remain fairly green. However, the price tag would have held customers back. Consumer electronics companies can gain from the FIS rules the insight of how much more customers are willing to pay for extra features and high greenness. There is little research of recommender system for green prod￾ucts. Liao (2008) proposed a green product recommendation sys￾tem in which the product database is queried against user designated weights of green criteria. EPEAT is used to represent the greenness of products as well. Liao’s system is effective in rec￾ommending green products; however, users are required to input the weights of certain green indices. In our system, on the other hand, users’ criteria of the three decision variables are implicitly gathered via users’ rating of products. For users who do not want the hassle, our system can use the quasi-conventional wisdom FIS to recommend product. Reducing the level of user intervention is important to the success of a recommender system. The pro￾posed architecture is in line with other commercially available rec￾ommender systems, such as Cinematch of Nextflix or Genius of iTunes, because in these systems, rating data is the minimal re￾quired input. 4. Conclusions In this paper, we proposed using recommender systems to aid the green shopping process and to promote green consumerism based upon the benefits of recommender systems and the FITD compliance technique. A proposed recommender system architec￾ture in the context of green consumer electronics was then de￾scribed and discussed. Previous research on this type of recommender system used weights of certain green indices as in￾puts. The architecture proposed in this paper accepts implicit and explicit criteria by modeling user preference and by considering ad hoc modification. The domain model is defined by the normalized Fig. 8. The rules of the quasi-conventional wisdom FIS. 9702 Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703

Y.-L Lee, F-H Huang/ Expert Systems with Applications 38(2011)9696-9703 scores of feature, price, and greenness, and the user model by an R-A.&Kao,C-Y.(2009) The architecture generates recommendations of consumer elec- Chen, Y-F.(2008). Herd behavior in purchasing books online. Computers in Human tronics to online customers by modeling their decision making Behavior,24(5).1977-1992 ith high-level decision variables. The distinctive part of the archi- Cooper, J.S,& Fava, J. A(2006). Life-cycle assessment practitioner survey tecture is the mutual bilateral relationship of either symbiosis or Freedman, J. L, Fraser, S C(1966) Compliance without pressure: The foot-in-the tibiosis between the pairs of price vs feature, price vs green ness, and feature vs greenness. User intervention is kept minimal, Gueguen, N.(2002) Foot-in-the-door technique. and computer-mediated nd on a par with commercially available recommender system aubl, G,& Rifts, V.(2000 Consumer decision making in online shopp The architecture has three types of recommendation Information environments: The effects of interactive decision aids. marketing science. filtering recommendation uses the Fis of a customer to sieve through the product database to find the products that are likely lijima.-& Ho, S(2007). Common structure and properties of filtering systems. to be preferable Candidate expansion takes the fuzzified variables Jang J-S.R(1993). ANFIS: Adaptive-network-based fuzzy inference systems.IEEE a reference product to find other comparable products. cro recommendation tells a customer what other customers have EA. (2001) Economics and electronic com nd directions for research. International Journal of electronic Commerce, 5(4 ught by tapping into the wisdom of like-minded crowd In add ion, the architecture uses a quasi-conventional wisdom FIs to rep- Kavcic. A(2004). Fuzzy user modeling for adaptation in educational hypermedia. resent new users whose rating data are nonexistent. In the future, we will expand the architecture to include other Liao y-o em In H-F types of green products since different product types have different Wang(Ed Web-based green products life cycle management systems:reverse ontology and eco-labeling schemes. In addition, more thorough Malcohn, E. Paulos. B. Stoeckle. A, Wang H H.-P(1994) Determinats of measures should be employed to ensure that user privacy and sys fectiveness for environmental certification and labeling programs. United States tem security are not compromised. The resultant user models have potential for green marketing, yet the data should remain anony Manouselis N stopoulou, C.(2007). Analysis and Classification of Multi mous if they are to be used for marketing purposes. Criteria Recommender Systems. World wide web: Internet and web information Acknowledgement e,1(1).9-48 Resnick, P. Varian, H (1997). Recommender systems. Communications of the ACM, The authors would like to express their gratitude to the mem- bers of the Green Supply Chain forum, of which the authors are Saaty, T(1987). Risk-its priority and probability: The analytic hierarchy process. also members, chaired by Professor Hsiao-Fan Wang of National Schein,, A. Popescul, A. Ungar, L. Pennock. D (2005). CROC: A new evaluation Tsing Hua University, and the members of the green Environment. Safety and Health Technology Division of Energy and Environment Shih, T K, Chiu, C-F. Hsu, H --& Lin, F(2002). An integrated framework for Research Laboratories(EEL)of Industrial Technology Research recommendation systems in e-commerce. Industrial Management 8 Dai Institute(ITRi). Their inputs are invaluable to us. This research is Systems.i02(8).417-431 traughan, R D,& Roberts, J. A(1999). Environmental segmentation alternatives: a funded by a grant from National Science Council of Taiwan under behavior in the new millennium Jourmal of Consumer the contract nsc 97-2221-E-324-018-MY3 takeing.16(6)558-575. Wan, Y. Menon, S,& Ramaprasad, A(2007). A classification of product comparison References Wang, P(2004 Recor Y. Zhang. A. Yao (Eds ) Computational we Adom generation of recomme ey of ossible extensions. IEEE nowledge and Data Engineering 17(6). 734-74 Zadeh, L(1965). Fuzzy sets. Information and ControL 8(3).338-353 nities and challenges. Industrial Zimmermann, H(2001). Fuzzy set and its applications. Kluwer Academic 8 Data Systems, 109(2)155-172 Cao, Y.& Li, Y(2007). An intelligent fuzzy-based recommendation system fc consumer electronic products. Expert Systems with Applications, 33(1). 230-240

scores of feature, price, and greenness, and the user model by an adaptive FIS. The architecture generates recommendations of consumer elec￾tronics to online customers by modeling their decision making with high-level decision variables. The distinctive part of the archi￾tecture is the mutual bilateral relationship of either symbiosis or antibiosis between the pairs of price vs. feature, price vs. green￾ness, and feature vs. greenness. User intervention is kept minimal, and on a par with commercially available recommender systems. The architecture has three types of recommendation. Information filtering recommendation uses the FIS of a customer to sieve through the product database to find the products that are likely to be preferable. Candidate expansion takes the fuzzified variables of a reference product to find other comparable products. Crowd recommendation tells a customer what other customers have bought by tapping into the wisdom of like-minded crowd. In addi￾tion, the architecture uses a quasi-conventional wisdom FIS to rep￾resent new users whose rating data are nonexistent. In the future, we will expand the architecture to include other types of green products since different product types have different ontology and eco-labeling schemes. In addition, more thorough measures should be employed to ensure that user privacy and sys￾tem security are not compromised. The resultant user models have potential for green marketing, yet the data should remain anony￾mous if they are to be used for marketing purposes. Acknowledgement The authors would like to express their gratitude to the mem￾bers of the Green Supply Chain forum, of which the authors are also members, chaired by Professor Hsiao-Fan Wang of National Tsing Hua University, and the members of the Green Environment, Safety and Health Technology Division of Energy and Environment Research Laboratories (EEL) of Industrial Technology Research Institute (ITRI). Their inputs are invaluable to us. This research is funded by a grant from National Science Council of Taiwan under the contract NSC 97-2221-E-324-018-MY3. References Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), 734–749. Bose, R. (2009). Advanced analytics: Opportunities and challenges. Industrial Management & Data Systems, 109(2), 155–172. Cao, Y., & Li, Y. (2007). An intelligent fuzzy-based recommendation system for consumer electronic products. Expert Systems with Applications, 33(1), 230–240. Chen, Y.-C., Shang, R.-A., & Kao, C.-Y. (2009). The effects of information overload on consumers’ subjective state towards buying decision in the internet shopping environment. Electronic Commerce Research and Applications, 8(1), 48–58. Chen, Y.-F. (2008). Herd behavior in purchasing books online. Computers in Human Behavior, 24(5), 1977–1992. Cooper, J. S., & Fava, J. A. (2006). Life-cycle assessment practitioner survey: Summary of results. Journal of Industrial Ecology, 10(4), 12–14. Freedman, J. L., & Fraser, S. C. (1966). Compliance without pressure: The foot-in-the￾door technique. Journal of Personality and Social Psychology, 4, 195–202. Guéguen, N. (2002). Foot-in-the-door technique and computer-mediated communication. Computers in Human Behavior, 18(1), 11–15. Häubl, G., & Trifts, V. (2000). Consumer decision making in online shopping environments: The effects of interactive decision aids. Marketing Science, 19(1), 4–21. Iijima, J., & Ho, S. (2007). Common structure and properties of filtering systems. Electronic Commerce Research and Applications, 6(2), 139–145. Jang, J.-S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference systems. IEEE Transactions on Systems, Man, and Cybernetics, 23(3), 665–685. Kauffman, R. J., & Walden, E. A. (2001). Economics and electronic commerce: Survey and directions for research. International Journal of Electronic Commerce, 5(4), 5–116. Kavcic, A. (2004). Fuzzy user modeling for adaptation in educational hypermedia. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, 34(4), 439–449. Liao, Y.-C. (2008). Green product retrieval and recommendations system. In H.-F. Wang (Ed.), Web-based green products life cycle management systems: reverse supply chain utilization (pp. 379–398). Idea Group Inc (IGI). Malcohn, E., Paulos, B., Stoeckle, A., & Wang, H. H.-P. (1994). Determinats of effectiveness for environmental certification and labeling programs. United States Environmental Protection Agency. Manouselis, N., & Costopoulou, C. (2007). Analysis and Classification of Multi￾Criteria Recommender Systems. World wide web: Internet and web information systems, 10(4), 415–441. Perny, P., & Zucker, J. (2001). Preference-based search and machine learning for collaborative filtering: The ‘‘Film-Conseil’’ movie recommender system. Information, Interaction, Intelligence, 1(1), 9–48. Resnick, P., & Varian, H. (1997). Recommender systems. Communications of the ACM, 40(3), 56–58. Saaty, T. (1987). Risk-its priority and probability: The analytic hierarchy process. Risk Analysis, 7(2), 159–172. Schein, A., Popescul, A., Ungar, L., & Pennock, D. (2005). CROC: A new evaluation criterion for recommender systems. Electronic Commerce Research, 5(1), 51–74. Shih, T. K., Chiu, C. -F., Hsu, H. -H., & Lin, F. (2002). An integrated framework for recommendation systems in e-commerce. Industrial Management & Data Systems, 102(8), 417–431. Straughan, R. D., & Roberts, J. A. (1999). Environmental segmentation alternatives: a look at green consumer behavior in the new millennium. Journal of Consumer Marketing, 16(6), 558–575. Wan, Y., Menon, S., & Ramaprasad, A. (2007). A classification of product comparison agents. Communications of the ACM, 50(8), 65–71. Wang, P. (2004). Recommendation based on personal preference. In Y. Zhang, A. Kandel, T. Lin, & Y. Yao (Eds.), Computational web intelligence: Intelligent technology for web applications (pp. 101–115). Singapore: World Scientific Publishing Company. Zadeh, L. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. Zimmermann, H. (2001). Fuzzy set theory – and its applications. Kluwer Academic Publishers. Y.-L. Lee, F.-H. Huang / Expert Systems with Applications 38 (2011) 9696–9703 9703