Study Of Recommendation Algorithms Based On Rating Elicitation

Recommender System has become one of the most promising techniques in the era of big data.It helps users quickly find items of interest from an overwhelming number of available items,usually in the form of providing the most useful items to users by a top-k ranking list.The tremendous growth of information as well as the number of users and products on the Internet add some key chanllenges to recommender systems.Oneof the chanllenges is the scalability issue.Collaborative filtering is the most successful technique in the area of recommender system.In order to improve the scalability of collaborative filtering recommender system,researchers proposed many cluster-based and parallel-based algorithms.Typically,in the phase of model training,they use all of the user behavior data without consider the quality of individual perference.And most existing methods are only applicable for the nerghborhood-based CF.However,from the perspective of the source dataset,we argue that not all user behavior data contribute to the final model equally,especially for active users who own plentiful behaviors.We propose that appropriately selected representative user preferences contain enough information to profile users,then generate sound recommendations in less time.Based on the above viewpoint,this paper first explore the relationship between the quantity of user preferences and recommendation performance by series of experiments,proposing the recommendation algorithm based on rating elicitation.Particularly,we consider both rating prediction and TopN recommendation task in this paper.Then,we propose a general rating elicitation framework and 3 rating elicitation strategies based on partitioning method,and 5 elicitation strategies based on aggregate statistics and information theory to select the most representative ratings for every user.Finally,we conduct abundant experiments on MovieLens and Netflix dataset with different elicitation strategies and experimentally conclude that part of representive user preferences are able to generate desirable recommendations at a lower computational cost,improving the scalability of the recommender system.And our proposed solution can apply for all collaborative filtering algorithms.