Effective recommendation methods by utilizing relationships among entities

Abstract

Recommendation can be made by either a single user or multi-users (i.e., a group). Data sparsity is a challenging problem in recommendation, which severely degrades performance of recommendation. Many previous works use auxiliary context information to resolve data sparsity. A heterogeneous information network (HIN) that consists of multiple types of nodes and links is frequently used as the context information since it has rich information between entities. In our dissertation, we propose recommendation methods for both a single user and a group by utilizing relationships among entities in HIN. In the first part, we propose a recommendation method for a single user by using topological features of HIN. Recommendation for a single user can be generalized to predict relationships between a user and an item. We predict existence of relationships between two entities by using structural characteristics and meaningful correlation. In the second part, we examine group recommendation considering group cohesion. Group cohesion denotes proportion of group members who actively participate in group activities. Since the group cohesion is an important factor in group recommendation, the concept of group cohesion needs to be properly considered. We also propose a hybrid model that reflects context information and rating data simultaneously. In the third part, we study dynamic group behavior modeling that utilizes context information. A novel group recommendation framework, namely DGC (short for Dynamic Group behavior modeling that utilizes Context information) is proposed. In DGC, we newly develop a transformation method that enables summarization of complex patterns in group decision making processes. To apply context information, we design a loss function based on semi-supervised learning that is combination of a supervised loss for label prediction and an unsupervised loss for context prediction. Experimental results show that our three methods provide significant performance improvements over other existing methods.