Challenging the Long Tail Recommendation on Heterogeneous Information Network

Recommender system, regarded as the lifeblood of many web systems, plays a critical role of discovering interested items from near-infinite inventory and exhibiting them to potential users. However, most of the existing recommender systems usually tend to recommend popular items and cannot discover niche items to surprise users, which is well known as the long tail problem. Data sparsity is the primary cause that users' historical data are not enough to learn their detail interests. Another reason is that the learning models have to neglect some individuality information for global optimum. In this paper, we propose a novel suite of heterogeneous information network (HIN) based methods for long tail recommendation. We first model both users' behavior data and context data with a unified HIN to handle the data sparsity issue. Then, we propose a basic solution that predict user's behavior based on its similar historical behaviors via Degree-aware General SimRank on HIN. To improve the accuracy, we investigate the contributions of different typed data, a novel enhancement framework is proposed based on deep neural network. Distinct from the traditional learning models, our methods predict user's behavior case by case which maximizes the personality information and can effectively discover the interested niche items. Experiments show that the proposed algorithm outperforms state-of-the-art techniques in long tail recommendation.