Recommender System Using Local Community Detection and Non-negative Matrix Factorization

A recommender system uses user(s) data to rank or predict recommendations and helps in discovering new products and services. Although collaborative filtering techniques are popularly used to provide recommendations, they are not immune to the cold start problem, a common challenge that arises when there is insufficient information about new items and users. On the other hand, communities discovered from a social network provide information about relationships among users, which is a good source of information to overcome the challenges in traditional recommendation systems. In this paper, we propose an approach that involves sampling the network for any seed node S using the personalized page rank algorithm. Then, by using Bethe Hessian matrix, we estimate the possible number of communities, and then by applying non-negative matrix factorization, we detect the overlapping communities in the social network. We define the belongingness factor for each community to know how much the user belongs to that community. We then use the detected communities for recommendations. Our approach effectively addresses the cold start problem and provides diverse recommendations. Experimental results on the MovieLens 1 M (user-item ratings) and Facebook Ego Network (user-user interactions) datasets demonstrate the viability of the proposed approach.