Constrained Social Community Recommendation

Abstract

In online social networks, users with similar interests tend to come together, forming social communities. Nowadays, user-defined communities become a prominent part of online social platforms as people who have joined such communities tend to be more active in social networks. Therefore, recommending explicit communities to users provides great potential to advance online services. In this paper, we focus on the constrained social community recommendation problem in real applications, where each user can only join at most one community. Previous attempts at community recommendation mostly adopt collaborative filtering approaches or random walk-based approaches, while ignoring social relationships between users as well as the local structure of each community. Therefore, they only derive an extremely sparse affinity matrix, which degrades the model performances. To tackle this issue, we propose ComRec which simultaneously captures both global and local information on the extended graph during pre-computation, speeding up the training process on real-world large graphs. In addition, we present a labeling component to improve the expressiveness of our framework. We conduct experiments on three Tencent mobile games to evaluate our proposed method. Extensive experimental results show that our ComRec consistently outperforms other competitors by up to 12.80% and 6.61% in the corresponding evaluation metrics of offline and online experiments, respectively.