A Study on Recommendation Ability in Collaborative Filtering by Fuzzy Co-Clustering with Exclusive Item Partition
Takaya Nakano, Katsuhiro Honda, Seiki Ubukata, Akira Notsu · 2016
Model-based collaborative filtering algorithms have been shown to be effective in constructing recommendation systems with fewer costs, and fuzzy co-clustering demonstrated a higher performance with 0-1 type purchase history data, where '0' elements can mean both not like and like but not bought yet. In this paper, the recommendation ability of a fuzzy co-clustering-based algorithm is further studied considering exclusive item partition in the co-clustering phase. In general co-clustering models, item typicality is independently estimated in each cocluster and all items can be shared by multiple clusters. Considering the specificity of some items, recommendation quality is expected to be improved in personalized recommendation. An experimental result demonstrates that introduction of exclusive constraints on specific items, which are preferred by only a small users, can contribute to improving recommendation ability in the recommendation phase.