Heter-Neighbor:a Neighborhood Based Recommendation Strategy In Heterogeneous Information Networks

With the rapid development of e-commerce systems, recommender systems have been widely used to identify user interest automatically and improve item sales, where neighborhood based models are the mostly used in real world applications. In traditional neighborhood models, predictions are made based on the effects of neighbors, which are identified by calculating user feedback based similarity matrix.The recent studies have shown that the combination of user-item feedback and external information, e.g., user social relations or movie actors, will improve the performance of recommender system as well as release the sufferings of time consuming calculation of traditional neighborhood models. To handle with increasing types of data, researchers proposed to use the models in academic heterogeneous information networks under recommender system context. However, most of these studies have not notice the difference between the uncertain user feedback and other types of relations, which could cause severe problems in performance as well as running time in real applications.In this paper, we analysis the difference in user activities between movie dataset and the academic dataset, based on which we redefine the heterogeneous information network in recommender systems (HIN-RECSYS) and suggest not identify initial neighborhoods by user-item relation to avoid time-consuming calculation similarity matrix. We propose a model called Heter-Neighbor based on HIN-RECSYS to find initial new neighborhoods of items and users only using external data and integrate them with SVD++model to optimize the feedback based similarity matrix, which could remain the advantages of traditional neighborhood models and avoid the increase on running time when importing new data. Empirical studies show that our approach outperforms both the SVD++and NMF based models.