Improving performances of Top-N recommendations with co-clustering method

Abstract

Listen

Translate article

Collaborative filtering has been widely used in many applications. The typical idea is to identify preferences of users by utilizing their interaction data over the whole items. However, this is sometimes inaccurate since users might share different preferences with different sets of items. In this paper, we put forward a new recommendation method based on collaborative filtering called User-Item Community Detection based Recommendation (UICDR) method. The new parameter-free and scalable community detection method is modified from our previous work. We derive a unipartite form of bipartite modularity and put forward a new network representation. By constructing a bipartite network with user-item interaction data, we first partitions users and items into several subgroups. After getting clusters with tightly linked users and items, traditional collaborative filtering models can be trained for each cluster. The results on four real-world data sets show that, the proposed UICDR significantly improves the performances of Top-N recommendations of several traditional collaborative filtering methods. In addition, UICDR is helpful to cold-start problem.