Fair sequential group recommendations

Fair Sequential Group Recommendations M A R I A S T R AT I G I , E VA G G E L I A P I T O U R A , J Y R K I N U M M E N M A A A N D K O S TA S S T E FA N I D I S HY562 Paper Presentation Γαυγιωτάκη Δέσποινα csdp1197 Recommender systems • Content-based method : recommended items are similar to other items that the user has already consumed in the past • Collaborative filtering method : a relevance function get items from similar enough users (ratings, feedback are used, textual review or like/dislike) produces the list of relevant items to the target user • Group recommendations: a group can make a query as well by applying a recommendation method to each member individually, and then aggregate the separate lists into one for the group. Group recommendation process Average approach : calculate the average score across all the group members’ preference scores for that item. In such a way, all members of the group are considered equals Least misery approach : use the minimum function rather than the average Unfairness in sequential group recommendations We want a system to recommend different items at each iteration, while retaining knowledge from past interactions, so as to keep the output diverse A system that has a memory and can adjust its recommendations accordingly If a group of friends requests a movie recommendation each week, both the average and least misery approaches fail through all iterations of the system. In the case of average, the outlier user is never satisfied, while in the case of least misery, the system recommends movies that are not highly relevant to anyone in the group. Contributions • Introduce the notion of sequential group recommendations : when adding the dimension of multiple iterations to typical group recommendation approaches, such as the average and least misery aggregation methods, the results do not ensure user satisfaction for all group members. • The concept of satisfaction : each member of the group has a degree of satisfaction for the items recommended at each iteration, as well as an overall satisfaction, gained by all the previous iterations of the group. • A sequential group recommendation model that takes into account the previous interactions of the group with the system and alters the influence that a group member has on the formation of the group recommendation list. • Experimentally show that the proposed model is superior to the standard group recommendation approaches Group recommendation aggregation methods Sequential group recommendations Introducing multiple rounds to the previous group recommendation process. Each group query to the system = iteration , is not a stand alone process, but a sequence of queries submitted to the system by the same group. Each iteration has the main components of a standard group recommendation method: single user recommendations followed by the aggregation phase This new model, in contrast to the previous one, introduces the notion of multiple iterations or sequence of recommendations. Target in altering the output of the recommender system, so that if a user was not satisfied in a previous iteration, potentially, he/she will be satisfied during the current one. Satisfaction Measure Gratification of each group member for the recommended items after each iteration of the system Single User Satisfaction : The degree of the satisfaction of each user in G to the group recommendation received at step j. Group Satisfaction : The satisfaction of the group G with respect to a group recommendation list is defined as the average of the satisfaction of the users in the group. Single User Satisfaction The quality of recommendations that the user receives as a member of the group compared to the quality of the recommendations that the user would have received as an individual . The overall satisfaction of user ui with respect to a sequence GR of µ iterations Group Satisfaction This measure indicates if the items reported to the group, are acceptable to its members Sequential Hybrid Aggregation Method Here, we can observe that a group member that was not satisfied in the previous iteration of the system, is satisfied in the next. A good example of this is User 4 where in the first iteration has a very low satisfaction score, and in the second has a higher one. This is a clear improvement over the results of the previous experiment, where User 4 was always the least satisfied member of the group // single user recommendation algorithm to each group member // members’ preference lists Gl // populate the set Algorithm //minimum satisfaction score of the group members in the previous iteration, from the maximum score Experimental Setup The Setup : ◦ Dataset ◦ 20M MovieLens Dataset : contains 20.000.263 ratings across 27.278 movies ◦ 138.493 users between January 09, 1995, and March 31, 2015 . ◦ chunks : each is added to the system, representing new information, and is used for locating the suggestions for the next iteration ◦ Group Formation ◦ ◦ ◦ ◦ the members of the group do not change between iterations, and all the members are present for each recommendation four types ofgroups, based on the similarity shared between the members (Pearson Correlation similarity measure , 0.7) 4 similar – 1 dissimilar , 3 similar – 2 dissimilar , 3 similar – 1 dissimilar – 1 dissimilar , 5 dissimilar For each group type, we generate 100 groups, and each user can only participate in one group per category Evaluation The evaluation ◦ Experimental Procedure ◦ F-score measure : harmonic mean of the groupSatO and groupDis measure ◦ To predict preference scores for a user, we use the Weighted Sum of Others Ratings , and take only the top100 most similar users to him/her. ◦ Recommend to the group the 10 items with the highest group preference score and assume that the system does not recommend previously recommended items Evaluation α Experiments : Τhe values that the variable α takes during the iterations Αverage α values of the 100 groups per type, per iteration The more diverse the group becomes, the more higher the values of α get because the more distinct opinions in the group are, the higher the disagreement between the user becomes, and thus the values of α become higher Evaluation The Group Type Experiments For all group types and all number of iterations, the method using dynamic α offers better results, than the rest of the methods. Even if the average method offers better group satisfaction, our proposed dynamic hybrid sequential method, more than makes up for it, with the far lower group disagreement • The overall group satisfaction for all group types decreases by the same degree because in the later iterations of the system the best items for the group have already been reported, and as stated we do not recommend items that have already been recommended in previous iterations. • When comparing the group satisfaction for the different values of α, we observe that the average method slightly outperforms the dynamic α. For 15 iterations, the average method offers a better overall group satisfaction by a slight factor, since it tries to offer the best results to the group as a whole. • The dynamic α, also clearly outperform the LM method, especially the more diverse the groups become. By far the worst performance is for α=1, since in each iteration of the system, we consider only one user, without taking into account the opinions of the rest of the group Even if the average method offers better group satisfaction, our proposed dynamic hybrid sequential method, more than makes up for it, with the far lower group disagreement Conclusion • Notion of sequential group recommendation • The standard group recommendation approaches are not favorable for sequential recommendations. • a minority opinion may be lost by the average method, • the preference of the group is determined by just one voice in the least misery • Sequential group recommendation model that takes into account the satisfaction of the members during the previous • Interactions of the group with the system.