myVU: A Next Generation Recommender system Based on observed Consumer Behavior and Interactive Evolutionary Algorithms Andreas Geyer-Schulz, Michael Hahsler, and Maximillian Jahn' Abteilung fur Informationswirtschaft Institut fur Informationsverarbeitung und Informationswirtschaft Abstract. my vU is a next generation recommender system based on observed consumer behavior and interactive evolutionary algorithms implementing customer relationship man- gement and one-to-one marketing in the educational and scientific broker system of a vir- tual university. my VU provides a personalized, adaptive www-based user interface for all mbers of a virtual university an scientific and educational Web-sites 1 Introduction In this article we describe my VU, a next generation recommender system based on observed consumer behavior and interactive evolutionary algorithms and report first results on its usage. my VU is a prototype of an educational and scientific rec- ommender system for providing routine recommendation and consulting services in mass universities to meet the challenge of reengineering the university posed in Tsichritzis(1999). my VU has been developed with the explicit goal of supporting life-long learning for large parts of the population and, therefore, must address the problems caused by the resulting high heterogeneity in students' background, capa bilities, and previous experience. In this article, however, we emphasize the role of evolution in the design of such systems and concentrate on the evolution strategies embedded into my VU. In section 2 we critically review existing interactive evo- lutionary algorithms and we identify the two main obstacles for their application to web-site personalization and design. Section 3 has the aim of explaining, ho evolutionary algorithms can be integrated in such a system: We treat a virtual uni- versity as an information market and base the evolutionary algorithm shaping the user interface on well-known statistics which are widely used in the retail industry namely ABC-analysis, market-basket analysis(see Blischok(1995))and consumer purchase histories(and -in the near future stochastic consumer choice models gner and Taudes(1987)). These statistics are computed from ap. propriately instrumented Web-server transaction logs. The evolutionary operators, namely fitness-biased selection and various kinds of mutation operators, influence the design and layout of the user-interface and thus influence user behavior. Ob usly, my VU can serve as a model of the business to customer interface of future e-commerce applications. my VU supports customer relationship management and
myVU: A Next Generation Recommender System Based on Observed Consumer Behavior and Interactive Evolutionary Algorithms Andreas Geyer-Schulz , Michael Hahsler , and Maximillian Jahn Abteilung fur¨ Informationswirtschaft, Institut fur¨ Informationsverarbeitung und Informationswirtschaft, Wirtschaftsuniversitat¨ Wien, A-1090 Wien, Austria Abstract. myVU is a next generation recommender system based on observed consumer behavior and interactive evolutionary algorithms implementing customer relationship management and one-to-one marketing in the educational and scientific broker system of a virtual university. myVU provides a personalized, adaptive WWW-based user interface for all members of a virtual university and it delivers routine recommendations for frequently used scientific and educational Web-sites. 1 Introduction In this article we describe myVU, a next generation recommender system based on observed consumer behavior and interactive evolutionary algorithms and report on first results on its usage. myVU is a prototype of an educational and scientific recommender system for providing routine recommendation and consulting services in mass universities to meet the challenge of reengineering the university posed in Tsichritzis (1999). myVU has been developed with the explicit goal of supporting life-long learning for large parts of the population and, therefore, must address the problems caused by the resulting high heterogeneity in students’ background, capabilities, and previous experience. In this article, however, we emphasize the role of evolution in the design of such systems and concentrate on the evolution strategies embedded into myVU. In section 2 we critically review existing interactive evolutionary algorithms and we identify the two main obstacles for their application to web-site personalization and design. Section 3 has the aim of explaining, how evolutionary algorithms can be integrated in such a system: We treat a virtual university as an information market and base the evolutionary algorithm shaping the user interface on well-known statistics which are widely used in the retail industry, namely ABC-analysis, market-basket analysis (see Blischok (1995)) and consumer purchase histories (and – in the near future – stochastic consumer choice models, for a survey see Wagner and Taudes (1987)). These statistics are computed from appropriately instrumented Web-server transaction logs. The evolutionary operators, namely fitness-biased selection and various kinds of mutation operators, influence the design and layout of the user-interface and thus influence user behavior. Obviously, myVU can serve as a model of the business to customer interface of future e-commerce applications. myVU supports customer relationship management and
Geyer-Schulz et al. one-to-one marketing as described in Kelly(1998)to a considerable extent. The evolutionary algorithm described above is common to all my VU recommender ser vices. All my VU recommender services available today(March 2000)are presented in section 4 together with a first analysis of their usage. Finally, we describe possible future improvements of the system 2 Interactive Evolutionary Algorithms for design In the opening lecture of the 1997 Genetic Programming Conference John Koza identified in his outlook on the future of evolutionary algorithms web-site design as one of the most promising(commercial) application areas. However, not much has been achieved in this area in the last three years. To be honest, we must be a little bit more precise. Interactive evolutionary and genetic algorithms have flourished in mu- sicandartsSeee.g.theweb-siteofCraigReynolds(http://www.red.com/cwr/evolve.html for examples on the www An evolutionary algorithm is characterized by a select-and-mutate approach with a population size of 1, a genetic algorithm in addition by a crossover opera- tor and a population size larger than 1. In this paper, however, we will neglect these differences. The reader may think of an evolutionary algorithm as a border line vari- ant of a genetic algorithm with zero probability of crossover and a population size 3 A survey of applications of interactive evolutionary and genetic algorithms for erspective is presented in Takagi(1996a)an akagi(1996b). However, we are not aware of any application of interactive evo- lutionary or genetic algorithms in the area of web-site personalization for e.g. cus- tomer relationship management and one-to-one marketing. Why Before we answer this question, we give a short review of the two generations of nteractive evolutionary and genetic algorithms used today by artists, musicians and engineers. The main feature of the first generation of these algorithms is that the user of such an algorithm is required to explicitely and directly assign a fitness value to each of the objects bred by the algorithm as shown in Figure 1. The ancestor of these algorithms is Richard Dawkins"Blind Watchmaker"which bred simple tree-based forms-so called biomorphs- with an interactive select-and-mutate approach. See Dawkins(1986). Smith's(1991)interactive genetic algorithm for breeding bug like creatures and Caldwell and Johnston's(1991)approach of assisting a witness in building a facial composite of a criminal suspect with an interactive genetic algo- rithm are two additional representatives of this class of algorithms. Combined with a virtual reality environment and bio-sensors as proposed by Hafner and RoBler (1995)interactive evolutionary al gorithms serve as advanced industrial design envi ronments The class of interactive evolutionary algorithms requires that we can phrase the problem as a search through some parameter space, that we can generate new can- didate solutions in near real-time and that the utility of those candidate solutions can easily be compared by humans, but not by means of a precisely defined fitness
2 Geyer-Schulz et al. one-to-one marketing as described in Kelly (1998) to a considerable extent. The evolutionary algorithm described above is common to all myVU recommender services. All myVU recommender services available today (March 2000) are presented in section 4 together with a first analysis of their usage. Finally, we describe possible future improvements of the system. 2 Interactive Evolutionary Algorithms for Design In the opening lecture of the 1997 Genetic Programming Conference John Koza identified in his outlook on the future of evolutionary algorithms web-site design as one of the most promising (commercial) application areas. However, not much has been achieved in this area in the last three years. To be honest, we must be a little bit more precise. Interactive evolutionary and genetic algorithms have flourished in music and arts. See e.g. the web-site of Craig Reynolds(http://www.red.com/cwr/evolve.html) for examples on the WWW. An evolutionary algorithm is characterized by a select-and-mutate approach with a population size of ✁ , a genetic algorithm in addition by a crossover operator and a population size larger than ✁ . In this paper, however, we will neglect these differences. The reader may think of an evolutionary algorithm as a border line variant of a genetic algorithm with zero probability of crossover and a population size of 1. A survey of applications of interactive evolutionary and genetic algorithms for system design from an engineering perspective is presented in Takagi (1996a) and Takagi (1996b). However, we are not aware of any application of interactive evolutionary or genetic algorithms in the area of web-site personalization for e.g. customer relationship management and one-to-one marketing. Why? Before we answer this question, we give a short review of the two generations of interactive evolutionary and genetic algorithms used today by artists, musicians and engineers. The main feature of the first generation of these algorithms is that the user of such an algorithm is required to explicitely and directly assign a fitness value to each of the objects bred by the algorithm as shown in Figure 1. The ancestor of these algorithms is Richard Dawkins’ “Blind Watchmaker” which bred simple tree-based forms – so called biomorphs – with an interactive select-and-mutate approach. See Dawkins (1986). Smith’s (1991) interactive genetic algorithm for breeding bug like creatures and Caldwell and Johnston’s (1991) approach of assisting a witness in building a facial composite of a criminal suspect with an interactive genetic algorithm are two additional representatives of this class of algorithms. Combined with a virtual reality environment and bio-sensors as proposed by Hafner ¨ and Roßler ¨ (1995) interactive evolutionary algorithms serve as advanced industrial design environments. The class of interactive evolutionary algorithms requires that we can phrase the problem as a search through some parameter space, that we can generate new candidate solutions in near real-time and that the utility of those candidate solutions can easily be compared by humans, but not by means of a precisely defined fitness
my VU: A Next Generation Recommender System 日团■ Interactive User assigns Fig. 1. Early interactive genetic algorithms function. Crucial for the success of this approach is that the number of candidate solutions that humans must evaluate is kept as low as possible, because baby-sitting for an interactive-evolutionary algorithm is not a task humans volunteer for This problem- the human fitness bottleneck- is addressed by interactive evo- lutionary algorithms of the second generation with the help of a meta-level genetic own in Figure 2 N 日■ Optimize object for a given fitness function fitness function! Fitness function GA Fig. 2. Interactive(meta) genetic algorithms. The interactive evolutionary algorithm at the meta-level breeds fitness functions which are used by a second genetic algorithm at the base -level to generate new can- didate solutions. For each fitness function, only the best solution is presented to the user for evaluation. The user evolves(either implicitely or explicitely) the fitness function which captures his taste or experience or aesthetics. An example of such an interactive algorithm, where the user explicitely manipulates a set of weights which express the relative importance of design factors as structural configuration harmony with the surrounding environment, and slenderness in the design of aes- thetic bridge structures, is presented in Furuta et al. (1996). Biles et al. (1996)carry the concept one step further. In GenJam, an interactive genetic algorithm for breed ing jazz-solos, the user implicitely chooses between fitness functions in the form of neural networks. By choosing a jazz-solo, he increases the fitness of the associated neural network of evolutionary algorithm which has been employed e.g. on the International netic Art II site by John Mount, Scott Neal Reilly and Michael Witbrock(described
myVU: A Next Generation Recommender System 3 Interactive GA User Population of objects sees breeds assigns fitness Fig. 1. Early interactive genetic algorithms. function. Crucial for the success of this approach is that the number of candidate solutions that humans must evaluate is kept as low as possible, because baby-sitting for an interactive-evolutionary algorithm is not a task humans volunteer for. This problem – the human fitness bottleneck – is addressed by interactive evolutionary algorithms of the second generation with the help of a meta-level genetic algorithm as shown in Figure 2. User sees Population GA Interactive Meta-GA Manipulate fitness function! select best Optimize object for a given fitness function! Fitness function Fig. 2. Interactive (meta) genetic algorithms. The interactive evolutionary algorithm at the meta-level breeds fitness functions which are used by a second genetic algorithm at the base-level to generate new candidate solutions. For each fitness function, only the best solution is presented to the user for evaluation. The user evolves (either implicitely or explicitely) the fitness function which captures his taste or experience or aesthetics. An example of such an interactive algorithm, where the user explicitely manipulates a set of weights which express the relative importance of design factors as structural configuration, harmony with the surrounding environment, and slenderness in the design of aesthetic bridge structures, is presented in Furuta et al. (1996). Biles et al. (1996) carry the concept one step further. In GenJam, an interactive genetic algorithm for breeding jazz-solos, the user implicitely chooses between fitness functions in the form of neural networks. By choosing a jazz-solo, he increases the fitness of the associated neural network. However, the human fitness bottleneck can be tackled with a different kind of evolutionary algorithm which has been employed e.g. on the International Genetic Art II site by John Mount, Scott Neal Reilly and Michael Witbrock (described
Geyer-Schulz et al. athttp://www.cs.cmu.edu/jmount/g3.html)atCarnegieMellon.Inthisapproach many users evaluate the objects bred by the genetic algorithm, the evaluations are collected and aggregated by the fitness function. A new generation of objects is generated as soon as a certain number of evaluations have been collected or a pre- specified period of time has passed. This approach suffers from two draw-backs first. the user receives no immediate feedback and he can not see. how his evalua tions have influenced the outcome of the algorithm and, second, if users have non- homogeneous utility functions and this is the case for heterogeneous user groups the averaging process inherent in the aggregation of the fitness values may lead to art objects whose aesthetic is a disappointment for all users. This effect is even more pronounced with small population sizes The main problems for applying interactive evolutionary algorithms currently ed by 1. The human fitness bottleneck caused by the required explicit evaluation of de- signs by the users 2. The heterogeneity of user groups which leads to non-homogeneous fitness func Bottleneck and Heterogeneous User Group Fitness 3 The Design of my VU: Addressing the humai To address the problems identified in the previous section, the design of my VU is based on the metaphor of an information market(broker) and on the principle of self-assessment of experience. This is an application of the economic principle of self-selection which has been suggested as a rationale for versioning information products by Shapiro and Varian(1999). In my VU we generate recommendations on information products(web-sites) from observed purchasing behavior for infor- mation products. The key idea is identify the information market as a system and the rest of the internet as its environment whenever the user follows an external link from the information market to an information product and crosses the system boundary, this is registered as a purchase incident. Use of internal links in the broker reveals the users' preferences for broker services. Information products in the virtual university have a rich meta-data description with one or more category attributes. In addition, in my VU users have the opportunity to incrementally establish a profile of their experience with each information product category visited(self-assessment of In the following the main loop of the interactive evolutionary algorithm of my VU shown in Figure 3 is explained in more detail. Numbers in parenthesis refer to arcs in Figure 3. In my VU recommendations are usually presented as ranked and labelled ists of suggested information products( web-sites) 1. The user perceives(1)a list of labelled links to other web-sites(e.g. his personal favorites shown in Figure 4)which constitutes an element of the user interface
4 Geyer-Schulz et al. at http://www.cs.cmu.edu/˜jmount/g3.html) at Carnegie Mellon. In this approach, many users evaluate the objects bred by the genetic algorithm, the evaluations are collected and aggregated by the fitness function. A new generation of objects is generated as soon as a certain number of evaluations have been collected or a prespecified period of time has passed. This approach suffers from two draw-backs: first, the user receives no immediate feedback and he can not see, how his evaluations have influenced the outcome of the algorithm and, second, if users have nonhomogeneous utility functions and this is the case for heterogeneous user groups, the averaging process inherent in the aggregation of the fitness values may lead to art objects whose aesthetic is a disappointment for all users. This effect is even more pronounced with small population sizes. The main problems for applying interactive evolutionary algorithms currently used by artists and engineers for personalized web-design are: 1. The human fitness bottleneck caused by the required explicit evaluation of designs by the users. 2. The heterogeneity of user groups which leads to non-homogeneous fitness functions. 3 The Design of myVU: Addressing the Human Fitness Bottleneck and Heterogeneous User Groups To address the problems identified in the previous section, the design of myVU is based on the metaphor of an information market (broker) and on the principle of self-assessment of experience. This is an application of the economic principle of self-selection which has been suggested as a rationale for versioning information products by Shapiro and Varian (1999). In myVU we generate recommendations on information products (web-sites) from observed purchasing behavior for information products. The key idea is identify the information market as a system and the rest of the Internet as its environment. Whenever the user follows an external link from the information market to an information product and crosses the system boundary, this is registered as a purchase incident. Use of internal links in the broker revealsthe users’ preferences for brokerservices. Information products in the virtual university have a rich meta-data description with one or more category attributes. In addition, in myVU users have the opportunity to incrementally establish a profile of their experience with each information product category visited (self-assessment of experience). In the following the main loop of the interactive evolutionary algorithm of myVU shown in Figure 3 is explained in more detail. Numbers in parenthesis refer to arcs in Figure 3. In myVU recommendations are usually presented as ranked and labelled lists of suggested information products (web-sites). 1. The user perceives(1) a list of labelled links to other web-sites (e.g. his personal favorites shown in Figure 4) which constitutes an element of the user interface
my VU: A Next Generation Recommender System Observatic Transaction Aggregation Agents Statistics Services Fig 3. my VU as interactive evolutionary algorithm 2. By following(2)one of these links the user reveals his preference for this web site. The oberservation agents record (4)the purchase incident in the transaction log. The fitness of an information product is proportional to the number of time it has been purchased. In my vU the fitness of an information product is thus not explicitely assigned by a user, it is computed from observed user behavior 3. By incrementally revealing their experience(novice, average, advanced or ex pert) with information product categories they have visited in previous ses- sions, my VU users establish(3)their own experience profile which is used for computing group specific recommendations(6)and for selecting the appropr ate group specific recommender service(7)for the user. This mechanism ad dresses the problem of heterogeneous user groups and has been recommended in Shapiro and Varian(1999) 4. Every night aggregation agents update the retail statistics with the purchase incidents(5)from the transaction log in accordance with the experience level of the user(6). Today, practitioners call this"web-mining. Such statistics include depending on the degree of anonymity chosen by the user) (a)If we can extract anonymous user sessions from the transaction log, market baskets can be analyzed. We compute for each information product yi the frequencies that, if information product yi is in a market basket, the other information products y1, .. i-1, yi+1,. yn are also in the basket. This requency is proportional to the conditional probability P(yiyi) to buy information product yi, if information product yi has been purchased in the same session (b)If we can extract pseudonymous user sessions from the transaction log, we can establish the purchase histories of users and combine these with their experience profiles. (Note, that pseudonymous means, that we do not know the true identity of a user, we only know that it is the same user. )We com- pute for all information products in a category the frequencies with which they are bought by a novice, an average user, an advanced user, and an
myVU: A Next Generation Recommender System 5 Observation Agents Experience Profile Recommendation Agents User (Fitness Function) Transaction Log Aggregation Agents Retail Recommender Statistics Services (Biased Selection) User Interface Discover Services (Mutation) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Fig. 3. myVU as interactive evolutionary algorithm. 2. By following (2) one of these links the user reveals his preference for this website. The oberservation agents record (4) the purchase incident in the transaction log. The fitness of an information product is proportional to the number of times it has been purchased. In myVU the fitness of an information product is thus not explicitely assigned by a user, it is computed from observed user behavior. 3. By incrementally revealing their experience (novice, average, advanced or expert) with information product categories they have visited in previous sessions, myVU users establish (3) their own experience profile which is used for computing group specific recommendations (6) and for selecting the appropriate group specific recommender service (7) for the user. This mechanism addresses the problem of heterogeneous user groups and has been recommended in Shapiro and Varian (1999). 4. Every night aggregation agents update the retail statistics with the purchase incidents (5) from the transaction log in accordance with the experience level of the user (6). Today, practitioners call this “web-mining”. Such statistics include (depending on the degree of anonymity chosen by the user): (a) If we can extract anonymous user sessions from the transaction log, market baskets can be analyzed. We compute for each information product ✂☎✄ the frequencies that, if information product ✂☎✄ is in a market basket, the other information products ✂✝✆✟✞✠✞✟✞✠✆ ✂✄☛✡ ✝✆ ✂✄✌☞ ✍✆✠✞✟✞✠✞ ✂✏✎ are also in the basket. This frequency is proportional to the conditional probability ✑✓✒✔✂✝✕✗✖ ✂✄✙✘ to buy information product ✂✝✕ , if information product ✂✄ has been purchased in the same session. (b) If we can extract pseudonymous user sessions from the transaction log, we can establish the purchase histories of users and combine these with their experience profiles. (Note, that pseudonymous means, that we do not know the true identity of a user, we only know that it is the same user.) We compute for all information products in a category the frequencies with which they are bought by a novice, an average user, an advanced user, and an
Geyer-Schulz et al. Netscape: my vU File Edit View Go Communicator Help 。3登应 N myvu Favorite Entries Anwendungsprojekt aus Ca :open in new window add to bookmarks others also use experts also use Ziel der Lehrveranstaltung ist die selbstandige Durchfuhrung eines kleinen, aber vollsfindigen EDU-Projekts, wind der Einsatz von Projektmanagementtechniken und softwareengineering Me thoden im Projekt trainiert French Universities open in new window add to bookmarks others also use Diplomanden-und Dissertantenseminar aus Informationswirtschaft open in new window add to bookmarks others also use experts also use rmafionswirts chaft mit besonderer Berucksichtigung der methodischen und nschaftstheoretischen Grundlagen zu usgewahlten actue len Themenbereichen des Faches, Fir Diplomanden und Dissertate Genetische Lemverfahren copen in new window add to bookmarks others also use experts also use RE VIED Lehrziel ist genethsche Lemuertehren(genehische Algonthhen, gene tische Programmierung und classifie Probleme, wie Lagerhaltung, Routenplanung scheduling, Prognoseverfahren, Informaton Retrieval, und fur strategische Entscheidungen einzusetzen. ③回 Fig. 4. Favorite Entries: A my VU-Service expert. Again, this is proportional to the conditional probability P(acec) that an information product ac from category c is purchased, if a user has experience level ec for category (c)For each user, the number of purchases of each information product in the relevant part of his purchase history. This statistic is used for ranking the entries of the Favorite Entries service shown in figure 4 (d) For each user, the number of purchases aggregated for information product categories (e)All products bought at least once by some user (f) From all transactions, an ABC-analysis of information products and cate- gories(for labelling the highest rated group with HOT) 5. As a user requires an information service of my vU, the recommendation agent responsible for this service generates the user interface element which results from this service. We distinguish between "recommender services which im
6 Geyer-Schulz et al. Fig. 4. Favorite Entries: A myVU-Service. expert. Again, this is proportional to the conditional probability ✑✓✒✔✚✜✛✝✖ ✢✍✛ ✘ that an information product ✚✣✛ from category ✤ is purchased, if a user has experience level ✢✝✛ for category ✤. (c) For each user, the number of purchases of each information product in the relevant part of his purchase history. This statistic is used for ranking the entries of the Favorite Entries service shown in figure 4. (d) For each user, the number of purchases aggregated for information product categories. (e) All products bought at least once by some user. (f) From all transactions, an ABC-analysis of information products and categories (for labelling the highest rated group with HOT). 5. As a user requires an information service of myVU, the recommendation agent responsible for this service generates the user interface element which results from this service. We distinguish between “recommender services” which im-
my VU: A Next Generation Recommender System plement fitness-proportional selection operators of the evolutionary algorithm and"discover services"which implement mutation operators of the evolution ary algorithm. Favorite Entries shown in Figure 4 is an example of such a rec- ommender service. Note, that a user interface element may consist of a bundle According to Holland(1975) fitness-biased selection operators serve the ex- ploitation of information. In my VU fitness-biased selection operators produce lists of information products ranked by their fitness. Ranking information products has a strong influence on the choice behavior of users. This fact has been established by several studies of user choice behavior in online environments. For a survey,see Introna and Nissenbaum(2000). In my vU, fitness-biased selection operators make repeat-buying behavior easier and thus increase user convenience. Convenience for repeat-buying has been identified as one of the key success factors for e-commerce sites by Bellmann et al. (1999) Mutation operators have the role of supporting the exploration of informatic my VU mutation operators either have the form of randomly generated link or category lists or of a randomly drawn banner leading to an information product in the virtual university. In the current implementation of my vU the mutation opera tors draw from the list of all information products ever purchased by some user and from the list of all information product categories ever used by some user. However leases of my VU additional mutation oper awing from other neigh- borhoods(e.g. all information products and categories in the virtual university, the conditional probability distribution of cross-selling,.)will be investigated. Muta- tion addresses an incentive problem of recommender systems discussed in Resnick and Varian(1997), namely, that users receiving recommendations diminish their search effort for information products and increasingly rely on a very narrow set of information products Consider, for example the Favorite Entries recommendation service illustrated in Figure 4. The link list is ranked according the the user's personal purchase fre quencies computed from his purchase history. Klicking on the"link"(the recom- mender service) with the label others also use in the line below the link Genetische Lernverfahren"(Genetic Machine Learning) leads to a list of informa- tion products y1, .. Un ranked according to the conditional probability P(y; l "Genetische Lernverfahren") The"link"(the recommender service)with the label: experts also use in- dicates that the user is an expert in the field of Genetic Machine Learning and it leads to a list of information products of the same category ranked according to the probability of being purchased by other experts for this category Recommendations based on the experience level of users for a category of in- formation products are only available for users who have revealed their experience for this category. This is a tit-for-tat strategy which adresses the free-riding problem inherent in recommender systems and it offers an incentive to the user to reveal his self-assessment of his experience for a category of information products. In the fu ture, we expect to exploit this information for learning progress monitoring and for team-building
myVU: A Next Generation Recommender System 7 plement fitness-proportional selection operators of the evolutionary algorithm and “discover services” which implement mutation operators of the evolutionary algorithm. Favorite Entries shown in Figure 4 is an example of such a recommender service. Note, that a user interface element may consist of a bundle of such services. According to Holland (1975) fitness-biased selection operators serve the exploitation of information. In myVU fitness-biased selection operators produce lists of information products ranked by their fitness. Ranking information products has a strong influence on the choice behavior of users. This fact has been established by several studies of user choice behavior in online environments. For a survey, see Introna and Nissenbaum (2000). In myVU, fitness-biased selection operators make repeat-buying behavior easier and thus increase user convenience. Convenience for repeat-buying has been identified as one of the key success factors for e-commerce sites by Bellmann et al. (1999). Mutation operators have the role of supporting the exploration of information. In myVU mutation operators either have the form of randomly generated link or category lists or of a randomly drawn banner leading to an information product in the virtual university. In the current implementation of myVU the mutation operators draw from the list of all information products ever purchased by some user and from the list of all information product categories ever used by some user. However, in future releases of myVU additional mutation operators drawing from other neighborhoods (e.g. all information products and categories in the virtual university, the conditional probability distribution of cross-selling, ...) will be investigated. Mutation addresses an incentive problem of recommender systems discussed in Resnick and Varian (1997), namely, that users receiving recommendations diminish their search effort for information products and increasingly rely on a very narrow set of information products. Consider, for example the Favorite Entries recommendation service illustrated in Figure 4. The link list is ranked according the the user’s personal purchase frequencies computed from his purchase history. Klicking on the “link” (the recommender service) with the label ::others also use in the line below the link “Genetische Lernverfahren” (Genetic Machine Learning) leads to a list of information products ✂✍✆✠✞✠✞✟✞✥✆ ✂✏✎ ranked according to the conditional probability ✑✓✒☛✂✍✕✗✖ “Genetische Lernverfahren”✘ . The “link” (the recommender service) with the label ::experts also use indicates that the user is an expert in the field of Genetic Machine Learning and it leads to a list of information products of the same category ranked according to the probability of being purchased by other experts for this category. Recommendations based on the experience level of users for a category of information products are only available for users who have revealed their experience for this category. This is a tit-for-tat strategy which adresses the free-riding problem inherent in recommender systems and it offers an incentive to the user to reveal his self-assessment of his experience for a category of information products. In the future, we expect to exploit this information for learning progress monitoring and for team-building
Geyer-Schulz et al. Netscape: myvu 吕回 File Edit View Go Communicator m myvu home What s new my vU bekommt ein neues Outfit! Die Testfassung ist nun offentlich zuganglich yvU Sie konnen mit der Voting Box abstimmen welche Fassung Ihnen am Besten gefallt logout Wie die zwische konnen Sie naturlich auch schon sehen e virtual university The virtual University L> personal services bookmarks newswire Tools Living Lectures Manuals fa WU Information VU Library entries Collection Recreation discover entries :discover categories personal information Search 100% ig. 5. my VU-Services. 4 The my vU Recommender Services Figure 5 shows the main page of the second version of my VU. The following rec- ommender services are available at the moment in my VU Favorite Entries(:: entries in the sidebar of Figure 5)have been shown in Figure 4 and discussed in the previous section. Favorite Categories (: categories in the sidebar of Figure 5)are a list of a user's most frequently used product categories. other also use in the context of a category is based on the conditional probability distribution that a user who buys a product from one category also buys a product from another he sidebar of Figure 5) list of randomly selected products from the virtual university for the purpose of stimulating the curiosity of my VU users
8 Geyer-Schulz et al. Fig. 5. myVU-Services. 4 The myVU Recommender Services Figure 5 shows the main page of the second version of myVU. The following recommender services are available at the moment in myVU: ✦ Favorite Entries (::entries in the sidebar of Figure 5) have been shown in Figure 4 and discussed in the previous section. ✦ Favorite Categories (::categories in the sidebar of Figure 5) are a list of a user’s most frequently used product categories. ::other also use in the context of a category is based on the conditional probability distribution that a user who buys a product from one category also buys a product from another category. ✦ Discover Entries (::discover entries in the sidebar of Figure 5) is a list of randomly selected products from the virtual university for the purpose of stimulating the curiosity of myVU users
my VU: A Next Generation Recommender System Service Usage(absolute) Usage(relative 628 3801% Newswire 4.66% Favorite entrie 21.19% Discover Entries 74 448% Discover Categ 478% Personal informatic 4.24% Table 1. Usage of my VU Services(January 26th, 2000-March 13th, 2000) Discover Categories(:: discover categories in the sidebar of Figure 5)is similar to Discover Entries, only at the level of information product cate- gories. The rationale for this service is to encourage the exploration of informa tion product categories which are new for a my VUuser They are grouped under the heading> favorites in the sidebar of the my V main page shown in Figure 5. Note, that recommender services are embedded as la bels into all my VU services including the my vU global bookmark service ( book- marks) which my VU users can access from wherever they are on the Internet In addition, the service:: recommender under the heading> customize allows the user to change his level of experience for categories he has visited in a previous session in the virtual university n Table I we have listed the actual usage of my VU services in the first Six weeks of operation(from January 26th, 2000 to March 13th, 2000). We see that all recommender services are actually used by my VU users. Favorite Entries is the most popular recommender services. It accounts for 20 percent of my VU service usage. Both mutation-based Discover services account for 4 percent of my VU usage each. Although this is only a very preliminary result, this seems to indicate that the mutation-based Discover services stimulate user curiosity and lead to an increase in the exploration efforts of users. Informally(e-mails and chat), student response to my VU has been favorable. At the end of March 2000 more than 200 users were 5 Future Research As usual, a lot remains to be done. From the perspective of evolutionary computa- tion which we have emphasized in this article, the following problems merit further What is a useful concept for a crossover operator for web-site design? A long term and more detailed study of the use of recommender services in
myVU: A Next Generation Recommender System 9 Service Usage (absolute) Usage (relative) Bookmarks 628 38.01% Newswire 77 4.66% Favorite Entries 350 21.19% Favorite Categories 228 13.8% Discover Entries 74 4.48% Discover Categories 79 4.78% Recommender Profile 146 8.84% Personal Information 70 4.24% Total 1652 100% Table 1. Usage of myVU Services (January 26th, 2000 – March 13th, 2000) ✦ Discover Categories (::discover categories in the sidebar of Figure 5) is similar to Discover Entries, only at the level of information product categories. The rationale for this service is to encourage the exploration of information product categories which are new for a myVU user. They are grouped under the heading > favorites in the sidebar of the myVU main page shown in Figure 5. Note, that recommender services are embedded as labels into all myVU servicesincluding the myVU global bookmark service (::bookmarks) which myVU users can access from wherever they are on the Internet. In addition, the service ::recommender under the heading > customize allows the user to change his level of experience for categories he has visited in a previous session in the virtual university. In Table 1 we have listed the actual usage of myVU services in the first six weeks of operation (from January 26th, 2000 to March 13th, 2000). We see that all recommender services are actually used by myVU users. Favorite Entries is the most popular recommender services. It accounts for 20 percent of myVU service usage. Both mutation-based Discover services account for 4 percent of myVU usage each. Although this is only a very preliminary result, this seems to indicate that the mutation-based Discover services stimulate user curiosity and lead to an increase in the exploration efforts of users. Informally (e-mails and chat), student response to myVU has been favorable. At the end of March 2000 more than 200 users were registered myVU users. 5 Future Research As usual, a lot remains to be done. From the perspective of evolutionary computation which we have emphasized in this article, the following problems merit further investigation: ✦ What is a useful concept for a crossover operator for web-site design? ✦ A long term and more detailed study of the use of recommender services in myVU is required
Geyer-Schulz et al. Several (obvious)refinements of the current services are not (yet) implemented cut-off values for truncating probability distributions, discounting strategies for older purchases, more flavors of mutation operators Appropriate bundling and naming of services Improved statistics of the dynamic development of the recommender system for an improved management of such systems A study of user acceptance. Are the changes in the my VU user interface unob trusive enough, so that the user accepts an adaptive user interface? What is the potential of combining recommendations based on behavior with content-related recommendations? 6 Links We invite the reader to visit my Vu (Http: //myvu. wu-wien. ac. at)at the Virtual Uni- versity(http://vu.wu-wien.ac.at)oftheViennaUniversityofEconomicsandBusi- ness Administration and to try the personalized recommender services described in his article 7 Acknowledgement We gratefully acknowledge the financial support of the Jubilaumsfonds of the Aus- trian National Bank under Grant No. 7925 without which my VU would never have een reall erences BELLMANN, S, LOHSE, G L, and JOHNSON, E.. (1999): Predictors of Online Buying BILES, J.A., ANDERSON PG o/ the ACM,42(12),32-38 P G, and LoGGl, L w(1996): Neural Network Fitness Func- tions for a Musical IGA. In: Procs. of the Int. ICSC Symp. on Intelligent Industrial Automation(A96)and Soff Computing ( SOC096, ICSC Academic Press, Reading, B39-B44. BLISCHOK, T.J. (1995): Every Transaction Tells a Story. Chain Store Age Executive, 71(3), CALDWELL, C and JOHNSTON, V.S.(1991): Tracking a Criminal Suspect through"Face- Space with a Genetic Algorithm. In:R K. Belew and L B. Booker(Eds ) Procs. of the 4th Int Conf on Genetic Algorithms(CGA91), Morgan Kaufmann, San Mateo 416-421 DAWKINS, R(1986): The Blind Watchmaker: Longman, Harlow. FURt A thetic desi, of ride states m icsi Application ot enein ergo ithm 415-421 HAFNER, U and ROSSLER, A (1995): Interactive Optimization with Evolutionary ies as a Tool for Virtual Reality. In: M. Freter(Ed ) Procs. Virtual Reality Wor IDG. Munchen. 369-378
10 Geyer-Schulz et al. ✦ Several (obvious) refinements of the current services are not (yet) implemented: cut-off values for truncating probability distributions, discounting strategies for older purchases, more flavors of mutation operators, ... ✦ Appropriate bundling and naming of services. ✦ Improved statistics of the dynamic development of the recommender system for an improved management of such systems. ✦ A study of user acceptance. Are the changes in the myVU user interface unobtrusive enough, so that the user accepts an adaptive user interface? ✦ What is the potential of combining recommendations based on behavior with content-related recommendations? 6 Links We invite the reader to visit myVU (http://myvu.wu-wien.ac.at) at the Virtual University (http://vu.wu-wien.ac.at) of the Vienna University of Economics and Business Administration and to try the personalized recommender services described in this article. 7 Acknowledgement We gratefully acknowledge the financial support of the Jubilaumsfonds ¨ of the Austrian National Bank under Grant No. 7925 without which myVU would never have been reality. References BELLMANN, S., LOHSE, G.L., and JOHNSON, E.J. (1999): Predictors of Online Buying Behavior. Communications of the ACM, 42(12), 32–38. BILES, J.A., ANDERSON, P.G., and LOGGI, L.W. (1996): Neural Network Fitness Functions for a Musical IGA. In: Procs. of the Int. ICSC Symp. on Intelligent Industrial Automation (IIA’96) and Soft Computing (SOCO’96), ICSC Academic Press, Reading, B39–B44. BLISCHOK, T.J. (1995): Every Transaction Tells a Story. Chain Store Age Executive, 71(3), 50–62. CALDWELL, C. and JOHNSTON, V.S. (1991): Tracking a Criminal Suspect through “FaceSpace” with a Genetic Algorithm. In: R.K. Belew and L.B. Booker (Eds.): Procs. of the 4th Int. Conf. on Genetic Algorithms (ICGA’91), Morgan Kaufmann, San Mateo, 416–421. DAWKINS, R. (1986): The Blind Watchmaker. Longman, Harlow. FURUTA, H., MAEDA, K., and WATANABE, E. (1995): Application of Genetic Algorithm to Aesthetic Design of Bridge Structures. Microcomputers in Civil Engineering, 10, 415–421. HAFNER, ¨ U. and ROSSLER, ¨ A. (1995): Interactive Optimization with Evolutionary Strategies as a Tool for Virtual Reality. In: M. Freter (Ed.): Procs. Virtual Reality World’95, IDG, Munchen, ¨ 369–378
