Collaborative recommendation based on social community detection

Abstract

Collaborative filtering algorithms have become one of the most used approaches to provide personalized services for users to deal with abundance of information. The traditional algorithms just use the explicit user-item rating matrix to find similar users or items. To improve the accuracy of the ratings predicted by the collaborative filtering algorithms, social information is widely incorporated into the traditional ones. Different with the existed works focus on directly connected neighbors, we consider the community between the users. We design the algorithms in two aspects: one is that the members in the same community have similar tastes and preferences, the other is that the member's taste is affected by the other members. We simplify these two factors as community similarity and community affection. Community similarity is incorporated into modifying the model-based collaborative filtering algorithm as the social community-based regularization (SCR), which improves 6.2% in mean absolute error (MAE) and 6.1% in root mean square error (RMSE) compared to the existed social recommendation algorithm. Community affection is incorporated into modifying the neighborhood-based collaborative filtering algorithm as the neighbor-based collaborative filtering based on community detection (NCFC) which improve 14.8% in MAE and 8.1% in RMSE compared to user-based collaborative filtering (UCF).