《电子商务 E-business》阅读文献：Managing uncertainty in group recommending processes

User Model User-Adap Inter(2009)19: 207-242 DOI10.1007/s11257-008-9061-1 ORIGINAL PAPER Managing uncertainty in group recommending Luis M. de Campos. Juan M. Fernandez.Luna Juan F Huete. Miguel A Rueda- Morales Received: 25 February 2008/ Accepted in revised form 26 October 2008/ Published online: 21 November 2008 O Springer Science+Business Media B V. 2008 Abstract While the problem of building recommender systems has attracted con siderable attention in recent years, most recommender systems are designed for rec- ommending items to individuals. The aim of this paper is to automatically recommend a ranked list of new items to a group of users. We will investigate the value of using Bayesian networks to represent the different uncertainties involved in a group rec ommending process, i.e. those uncertainties related to mechanisms that govern the interactions between group members and the processes leading to the final choice or recommendation. We will also show how the most common aggregation strategies might be encoded using a Bayesian network formalism. The proposed model can be considered as a collaborative Bayesian network-based group recommender systen where group ratings are computed from the past voting patterns of other users with Probabilistic hical models L M. de Campos.J M. Fernandez-Luna.J F Huete(). M. A. Rueda-Moral Department of Computer Science and Artificial Intelligence, University of Granada, 8071 Granada, Spain e-mail: Ici(@desai.ugres e-mail: jmfluna@desai. ugres M. A. Rueda- Morales e-mail: mrueda (@desai. ugres

User Model User-Adap Inter (2009) 19:207–242 DOI 10.1007/s11257-008-9061-1 ORIGINAL PAPER Managing uncertainty in group recommending processes Luis M. de Campos · Juan M. Fernández-Luna · Juan F. Huete · Miguel A. Rueda-Morales Received: 25 February 2008 / Accepted in revised form : 26 October 2008 / Published online: 21 November 2008 © Springer Science+Business Media B.V. 2008 Abstract While the problem of building recommender systems has attracted con￾siderable attention in recent years, most recommender systems are designed for rec￾ommending items to individuals. The aim of this paper is to automatically recommend a ranked list of new items to a group of users. We will investigate the value of using Bayesian networks to represent the different uncertainties involved in a group rec￾ommending process, i.e. those uncertainties related to mechanisms that govern the interactions between group members and the processes leading to the final choice or recommendation. We will also show how the most common aggregation strategies might be encoded using a Bayesian network formalism. The proposed model can be considered as a collaborative Bayesian network-based group recommender system, where group ratings are computed from the past voting patterns of other users with similar tastes. Keywords Group recommending · Management of uncertainty · Probabilistic graphical models L. M. de Campos · J. M. Fernández-Luna · J. F. Huete (B) · M. A. Rueda-Morales Department of Computer Science and Artificial Intelligence, University of Granada, 18071 Granada, Spain e-mail: jhg@decsai.ugr.es L. M. de Campos e-mail: lci@decsai.ugr.es J. M. Fernández-Luna e-mail: jmfluna@decsai.ugr.es M. A. Rueda-Morales e-mail: mrueda@decsai.ugr.es 123

L. M. de Campos et al. 1 Introduction Recommender systems (RS) provide specific suggestions about items(or actions) within a given domain which may be considered of interest to the user(Resnick and Varian 1997). Depending on the information used when recommending, traditional RS are mainly classified into content and collaborative-based RS, although hybrid approaches do exist. The first type recommends a product by considering its content similarity with those products in which the user has previously expressed an interest. The second alternative attempts to identify groups of people with similar tastes to the user and to recommend items that they have liked Most RS are designed for individual use, i.e. there is an active user who receives recommendations about certain products once they have logged on to the system In this paper, we will focus on the related problem of group recommending(Gr) here the objective is to obtain recommendations for groups of people(Jameson and Smyth 2007). This kind of Rs is appropriate for domains where a group of people participate in a single activity such as watching a movie or going on vacation and also in situations where a single person must make a decision on behalf of the group. In one way or another, GR involves merging different individual preferences. In these situations, it is natural that one of the most important issues is the search for an aggregation mechanism to obtain recommendations for the group. According to Pennock and Wellman(2005)". there is nothing close to a single well-accepted normative basis for group beliefs, group preferences or group decision making. and many aggregation strategies can therefore be found in literature for group decisions (Masthoff 2004; Masthoff and Gatt 2006; Yu et al. 2006; Jameson and Smyth 2007) It is typically assumed that member preferences are given using a rating domain (let us say from 5*, really like, to 1*, really hate). An aggregation strategy is then used to determine the group rating. For example, let us consider a group with three individuals, John, Ann and Mary, where John rates a product 5, Ann rates it 2, and Mary rates it 5". Following an average aggregation criterion, we could then say that the group duct is 4 As in the previous example, the methods proposed in GR literature(see Jameson and Smyth 2007 for a review) do not deal with uncertainty. They assume that the inputs of the aggregation functions (i.e. user preferences)are precise and use a merging strat egy to compute precise outputs. This assumption is not necessarily true, especially if we consider that the user's preferences are normally determined by means of auto matic mechanisms In these cases, a probability distribution over the candidate ratings might be used to express user likelihoods. For example, Table I shows the probability distributions representing the preferences of three users(A, B, and C). In this case Table 1 User ratings for a given User

208 L. M. de Campos et al. 1 Introduction Recommender systems (RS) provide specific suggestions about items (or actions) within a given domain which may be considered of interest to the user (Resnick and Varian 1997). Depending on the information used when recommending, traditional RS are mainly classified into content and collaborative-based RS, although hybrid approaches do exist. The first type recommends a product by considering its content similarity with those products in which the user has previously expressed an interest. The second alternative attempts to identify groups of people with similar tastes to the user and to recommend items that they have liked. Most RS are designed for individual use, i.e. there is an active user who receives recommendations about certain products once they have logged on to the system. In this paper, we will focus on the related problem of group recommending (GR), where the objective is to obtain recommendations for groups of people (Jameson and Smyth 2007). This kind of RS is appropriate for domains where a group of people participate in a single activity such as watching a movie or going on vacation and also in situations where a single person must make a decision on behalf of the group. In one way or another, GR involves merging different individual preferences. In these situations, it is natural that one of the most important issues is the search for an aggregation mechanism to obtain recommendations for the group. According to Pennock and Wellman (2005) “... there is nothing close to a single well-accepted normative basis for group beliefs, group preferences or group decision making.”, and many aggregation strategies can therefore be found in literature for group decisions (Masthoff 2004; Masthoff and Gatt 2006; Yu et al. 2006; Jameson and Smyth 2007). It is typically assumed that member preferences are given using a rating domain (let us say from 5∗, really like, to 1∗, really hate). An aggregation strategy is then used to determine the group rating. For example, let us consider a group with three individuals, John, Ann and Mary, where John rates a product 5∗, Ann rates it 2∗, and Mary rates it 5∗. Following an average aggregation criterion, we could then say that the group rating for this product is 4∗. As in the previous example, the methods proposed in GR literature (see Jameson and Smyth 2007 for a review) do not deal with uncertainty. They assume that the inputs of the aggregation functions (i.e. user preferences) are precise and use a merging strat￾egy to compute precise outputs. This assumption is not necessarily true, especially if we consider that the user’s preferences are normally determined by means of auto￾matic mechanisms. In these cases, a probability distribution over the candidate ratings might be used to express user likelihoods. For example, Table 1 shows the probability distributions representing the preferences of three users (A, B, and C). In this case, Table 1 User ratings for a given item User 1* 2* 3* 4* 5* A 0.2 0.2 0.2 0.19 0.21 B 0 000.1 0.9 C 0.49 0 0 0 0.51 123

Uncertainty in group recommending lthough 5" might be considered the most probable rating, we will not have the same confidence about every situation. Surprisingly, little attention has been paid in GR literature to the problem of man aging uncertainty although it has been well established in the general group decision framework(see Clemen and winkler 1999: Genest and Zidek 1986 for a review ) In his paper, therefore, we will focus on this particular problem. We maintain that two different sources of uncertainty can be found in group recommending processes: the uncertainty shown when user preferences are set, i.e. the user's personal opinion about an item or feature; and the uncertainty which is inherent to the merging process The purpose of this paper is to investigate the value of using Bayesian networks (BN) to represent how different individuals in a group interact in order to make a final choice or recommendation. In our approach, the BN formalism is used to represent both the interactions between group members and the processes leading to the final choice or recommendation We will show how common decision rules in literature could be managed by adequately designing canonical models with the BN language, thereby shedding new light on the combination processes. Discussion about subjects such as how the groups are formed, how long they have existed, relationships between group members, how the group might interact to reach a consensus, etc are beyond the scope of this paper. We shall assume that all the individuals use the same set of labels to express their preferences for an item, and that these preferences are represented by means of a probability distribution(probably estimated from a data set) We consider BNs appropriate because they combine a qualitative representation of the problem through an explicit representation of the dependence relationships between items, users and groups, with a quantitative representation by means of a set of probability distributions to measure the strength of these relationships. Throughout the process, we must consider the computational aspects of the RS, where the sparse ness of the data and the fact that the ranking should be computed in real time represent two challenges. The second section of this paper briefly examines group recommender systems nd related work. Section 3 presents the proposed BN-based model which enables the interaction between individuals to be represented. Section 4 examines how to represent the strength of the individuals'interactions (i.e. conditional probability distributions) and Sect. 5 discusses how inference is performed in order to make recommendations to the group. Section 6 examines the experimental framework. Section 7 discusses e experimental results obtained when considering uncertainty in individual ratings and in Sect. 8 we study those situations where the process behind the group rating is also uncertain. Finally, our conclusions and comments regarding further research are discussed in Sect. 9 2 Classification of group recommender systems and related work Although GR is quite a new research topic, many papers on this problem have already been published. The specific objectives of recommender systems in the research pub- lished so far are determined by the characteristics of the domain for which the system has been developed. These characteristics significantly affect the choice of design and

Uncertainty in group recommending 209 although 5∗ might be considered the most probable rating, we will not have the same confidence about every situation. Surprisingly, little attention has been paid in GR literature to the problem of man￾aging uncertainty although it has been well established in the general group decision framework (see Clemen and Winkler 1999; Genest and Zidek 1986 for a review). In this paper, therefore, we will focus on this particular problem. We maintain that two different sources of uncertainty can be found in group recommending processes: the uncertainty shown when user preferences are set, i.e. the user’s personal opinion about an item or feature; and the uncertainty which is inherent to the merging process. The purpose of this paper is to investigate the value of using Bayesian networks (BN) to represent how different individuals in a group interact in order to make a final choice or recommendation. In our approach, the BN formalism is used to represent both the interactions between group members and the processes leading to the final choice or recommendation. We will show how common decision rules in literature could be managed by adequately designing canonical models with the BN language, thereby shedding new light on the combination processes. Discussion about subjects such as how the groups are formed, how long they have existed, relationships between group members, how the group might interact to reach a consensus, etc. are beyond the scope of this paper. We shall assume that all the individuals use the same set of labels to express their preferences for an item, and that these preferences are represented by means of a probability distribution (probably estimated from a data set). We consider BNs appropriate because they combine a qualitative representation of the problem through an explicit representation of the dependence relationships between items, users and groups, with a quantitative representation by means of a set of probability distributions to measure the strength of these relationships. Throughout the process, we must consider the computational aspects of the RS, where the sparse￾ness of the data and the fact that the ranking should be computed in real time represent two challenges. The second section of this paper briefly examines group recommender systems and related work. Section 3 presents the proposed BN-based model which enables the interaction between individuals to be represented. Section 4 examines how to represent the strength of the individuals’ interactions (i.e. conditional probability distributions) and Sect. 5 discusses how inference is performed in order to make recommendations to the group. Section 6 examines the experimental framework. Section 7 discusses the experimental results obtained when considering uncertainty in individual ratings and in Sect. 8 we study those situations where the process behind the group rating is also uncertain. Finally, our conclusions and comments regarding further research are discussed in Sect. 9. 2 Classification of group recommender systems and related work Although GR is quite a new research topic, many papers on this problem have already been published. The specific objectives of recommender systems in the research pub￾lished so far are determined by the characteristics of the domain for which the system has been developed. These characteristics significantly affect the choice of design and 123

L. M. de Campos et al. RECOMM STRATEGIES A. Prof INFORMATION SOURCE Fig. 1 Classification of Group Recommending Systems each publication therefore focuses on a specific issue( from how to acquire information about group preferences or how the system generates and explains the recommenda- tions to studying the mechanism used to reach a consensus (Jameson and Smyth 2007)). As a result, relating the different approaches is a difficult task. In this section, we will present a new classification taxonomy for group recom ending systems. This classification is based on three independent components of primary importance in the design of a group recommending system and not on the particular techniques used to solve each problem: the information source, the aggre gation criterion used to make the recommendations, and the user's interaction with the system Figure I shows a graphical representation of the proposed classification Source of information: This classification criterion, which has been borrowed from classical RS literature(Adomavicius and Tuzhilin 2005), distinguishes betweer content-based(CB)and collaborative filtering(CF). In the first case, the recom- mended items are those which are similar to the ones that individuals have found interesting in the past. As a result, it is necessary to analyze the content's features recommending. The second alternative considers that the recommendations for a target product have been obtained by considering how people with similar tastes rated a product in the past. These systems are based on the idea that people will agree in future evalua- tions if they have also agreed in their past evaluations. The information sources are therefore the preference ratings given by similar users A new category can obviously be obtained if we consider hybrid approaches that combine both(collaborative and content-based) methods Recommendation strategies: Once we have the information to hand, the strategy used for aggregating this infor mation is a central point in group recommending, and generally in any group deci sion process. In this case, two different approaches can be distinguished. The first I Without loss of generality, we have decided not toinclude thi ory in our taxonomy since, to the best of our knowledge, no study has tried to combine both techniques

210 L. M. de Campos et al. USER active passive INFORMATION SOURCE content collaborative A. Rec. A. Prof. RECOMM. STRATEGIES Fig. 1 Classification of Group Recommending Systems each publication therefore focuses on a specific issue (from how to acquire information about group preferences or how the system generates and explains the recommenda￾tions to studying the mechanism used to reach a consensus (Jameson and Smyth 2007)). As a result, relating the different approaches is a difficult task. In this section, we will present a new classification taxonomy for group recom￾mending systems. This classification is based on three independent components of primary importance in the design of a group recommending system and not on the particular techniques used to solve each problem: the information source, the aggre￾gation criterion used to make the recommendations, and the user’s interaction with the system. Figure 1 shows a graphical representation of the proposed classification. – Source of information: This classification criterion, which has been borrowed from classical RS literature (Adomavicius and Tuzhilin 2005), distinguishes between content-based (CB) and collaborative filtering (CF). In the first case, the recom￾mended items are those which are similar to the ones that individuals have found interesting in the past. As a result, it is necessary to analyze the content’s features for recommending. The second alternative considers that the recommendations for a target product have been obtained by considering how people with similar tastes rated a product in the past. These systems are based on the idea that people will agree in future evalua￾tions if they have also agreed in their past evaluations. The information sources are therefore the preference ratings given by similar users. A new category can obviously be obtained if we consider hybrid approaches that combine both (collaborative and content-based) methods.1 – Recommendation strategies: Once we have the information to hand, the strategy used for aggregating this infor￾mation is a central point in group recommending, and generally in any group deci￾sion process. In this case, two different approaches can be distinguished. The first 1 Without loss of generality, we have decided not to include this category in our taxonomy since, to the best of our knowledge, no study has tried to combine both techniques in the group recommending framework. 123

approach, aggregating recommendations(Ar), is a two-step strategy, where an indi- vidual recommendation is first obtained for each group member, and then a common recommendation is obtained by merging these individual recommendations. In the second approach, aggregating profiles(AP), the objective is to obtain a common profile by representing group preferences. This can be done explicitly, where the ndividuals use a common group account to give their preferences, or implicitly by means of an aggregation mechanism for the different individuals' profiles or Individual interactions Finally, a group recommending system can also be categorized by considering the way in which the users interact with the system. The individuals can be dichoto- mized into passive members(PM)and active members(AM). Focusing on the active embers, the final purpose is to reach a consensus between the group members and like many decision support system approaches, it is necessary for the users to eval- uate the system recommendations. In contrast, when the members are passive, the final purpose is only to provide a recommendation to the group, as might be the case when using an RS in a marketing campaign. In this situation, the individuals do not interact with the system in order to evaluate the proposed recommendations Since we use three non-overlapping criteria for classification purposes, a given GrS can be classified using three labels, one for each category. For instance, a GRS can be classified as CB-+AP+PM if the group profile is obtained by combining the infor mation about the content of the items which have been previously evaluated by each user. This profile will be used to send the final recommendations to the group 2.1 Related work Once the taxonomy has been presented, we will then go on to classify previously published GR systems CB+AP+PM: most published GRSs might be included in this category. For exam ple, let us consider MusicFX (McCarthy and Anagnost 2000). Given a database of member preferences for musical genres(each user rates each of the 91 genres on a five-point scale), the group profile is computed by summing the squared individual preferences. Using a weighted random selection operator, the next music station to be played is then selected. No interaction with the system is possible except by changing user preferences The inputs in the case of group modeling(Masthoff 2004)are user preferences(rat- ings)for a series of programs, and in this paper we study the performance of several aggregation strategies. The article(Yu et al. 2006) presents various TV program recommendations for multiple viewers by merging individual user preferences on eatures(e.g. genre, actor, etc. )to construct a group profile. The aim of the aggrega tion strategy is to minimize the total distance in such a way that the merged profile s close to most user preferences, thereby satisfying most of the group. CB+AP+AM: The Travel Decision Forum (Jameson 2004)was developed to help a group of users agree on the desired attributes of a vacation. This system allows

Uncertainty in group recommending 211 approach, aggregating recommendations(AR), is a two-step strategy, where an indi￾vidual recommendation is first obtained for each group member, and then a common recommendation is obtained by merging these individual recommendations. In the second approach, aggregating profiles (AP), the objective is to obtain a common profile by representing group preferences. This can be done explicitly, where the individuals use a common group account to give their preferences, or implicitly, by means of an aggregation mechanism for the different individuals’ profiles or preferences. – Individual interactions Finally, a group recommending system can also be categorized by considering the way in which the users interact with the system. The individuals can be dichoto￾mized into passive members(PM) and active members(AM). Focusing on the active members, the final purpose is to reach a consensus between the group members and, like many decision support system approaches, it is necessary for the users to eval￾uate the system recommendations. In contrast, when the members are passive, the final purpose is only to provide a recommendation to the group, as might be the case when using an RS in a marketing campaign. In this situation, the individuals do not interact with the system in order to evaluate the proposed recommendations. Since we use three non-overlapping criteria for classification purposes, a given GRS can be classified using three labels, one for each category. For instance, a GRS can be classified as CB+AP+PM if the group profile is obtained by combining the infor￾mation about the content of the items which have been previously evaluated by each user. This profile will be used to send the final recommendations to the group. 2.1 Related work Once the taxonomy has been presented, we will then go on to classify previously published GR systems. – CB+AP+PM: most published GRSs might be included in this category. For exam￾ple, let us consider MusicFX (McCarthy and Anagnost 2000). Given a database of member preferences for musical genres (each user rates each of the 91 genres on a five-point scale), the group profile is computed by summing the squared individual preferences. Using a weighted random selection operator, the next music station to be played is then selected. No interaction with the system is possible except by changing user preferences. The inputs in the case of group modeling (Masthoff 2004) are user preferences (rat￾ings) for a series of programs, and in this paper we study the performance of several aggregation strategies. The article (Yu et al. 2006) presents various TV program recommendations for multiple viewers by merging individual user preferences on features (e.g. genre, actor, etc.) to construct a group profile. The aim of the aggrega￾tion strategy is to minimize the total distance in such a way that the merged profile is close to most user preferences, thereby satisfying most of the group. – CB+AP+AM: The Travel Decision Forum (Jameson 2004) was developed to help a group of users agree on the desired attributes of a vacation. This system allows 123