Individual Diversity Preference Aware Neural Collaborative Filtering

Abstract

The diversified recommendation of recommender systems enriches user experiences by diversifying recommendation lists. However, the conventional post-processing strategy, which re-ranking the recommendation lists by diversity measurement, could cause a trade-off between accuracy and diversity. An improvement of one performance usually sacrifices the other one. Thus, we propose an in-processing method named Individual Diversity Preference Aware Neural Collaborative Filtering that takes accuracy and diversity together as optimal direction during the recommendation process. It consists of two filtering processes: collaborative filtering and diversity filtering. The items with high accuracy and high diversity to users could be on the recommendation lists. Besides, to meet different diverse needs of users, we model the diversity preference of an individual user by the items he/she has interacted with. As the front preparation of our method, we pre-train the item embeddings by applying a graph embedding technique on the item-linking knowledge graph. Since the pre-training is independent of recommendation, the pre-trained item embeddings could be used for objective diversity comparison among different recommendation models. To validate our proposed method, we conduct extensive experiments on Movielens datasets. Experimental results suggest that our approach has achieved high accuracy and high diversity of recommendations.