Collaborative Filtering Based On User-item Pairwise Blocking

To our knowledge, almost all existing collaborative filtering need to find elusive neighbouring relationship between users or between items based on some similarity measure in some space. However, a hypothesis behind most previous works is that neighbouring or similar relationship of users would be static across the whole set of items, which is not true in the reality. Two user who share similar taste on some items may have totally different opinion on others. Users maybe clustering into many groups in terms of their opinion on a set of items, these groups would collapse and new users’cluster structure would be built in terms of their opinion on a new set of items. Analogously, clusters of items formed according to their popularity among a group of users would be disintegrated when encounter a new group of users. In a nut-shell, users’cluster structure would vary across the set of items, and vice versa, clus-ter’structure of items would vary across the set of users. In this paper, we strive to find block structure embedded in the ratings matrix. We use block structure to build the clusters of users and of items and characterize the interaction between users and items. Every block consists of many users and items, all users involved in some block share similar opinions on all of the items in the same block, on the other hand, in the same block, all items share similar popularity among the users. Every block is an mean-ingful colony resided by users and items, which reflecting the extent of coalescent of users’taste with the latent attributes of items and the interaction between the groups of users and of items. we suggest a general framework for collaborative filtering based on user-item pairwise blocking, which involves the unified feature extraction of users and items, the alignment of users to items, and clustering the item with the alignment representation, so to obtain the block structure we anticipate. At last, we adopt existing collaborative to learn the latent factor on the global level and on the block level. We de-velop a implementation of this framework and experimental evidences show that, with utilization of these taste blocks, performance of recommendation can be improved.