Collaborative filtering (CF) based methods have become the most popular techniques for recommender systems (RSs). In recent years, various types of side information such as social connections among users and metadata of items have been introduced into CF and shown to be effective for improving recommending performance. Moreover, side information helps to alleviate data sparsity and cold start problems of conventional CF based RSs. However, previous works process different types of information separately, thus losing information that might exist across different types of side information.

In this thesis, we explore methods to enhance RS with various side information. We start with the incorporation of an important type of side information, i.e., social connections among users, i...[ Read more ]

Collaborative filtering (CF) based methods have become the most popular techniques for recommender systems (RSs). In recent years, various types of side information such as social connections among users and metadata of items have been introduced into CF and shown to be effective for improving recommending performance. Moreover, side information helps to alleviate data sparsity and cold start problems of conventional CF based RSs. However, previous works process different types of information separately, thus losing information that might exist across different types of side information.

In this thesis, we explore methods to enhance RS with various side information. We start with the incorporation of an important type of side information, i.e., social connections among users, into a state-of-the-art matrix factorization (MF) method, namely, Local LOw Rank Matrix Approximation (LLORMA). We propose our Social LOcal Matrix Approximation (SLOMA), which exploits social relationship in decomposing the user-item matrix into low-rank matrices. Experimental results obtained from two real-world datasets demonstrate the superiority of SLOMA to LLORMA in the rating prediction task.

Next, we study the application of Heterogeneous Information Network (HIN) to enhance CF based recommendation methods. HIN is a flexible scheme for representing the connections between different types of information. Since HIN could be a complex graph representing multiple types of relations between entity types, we need to tackle two challenging issues facing HIN-based RSs: How to capture the complex semantics determining the similarities between users and items in a HIN, and how to fuse the heterogeneous information to support recommendation. We propose to apply metagraph to HIN-based RSs to overcome the former problem and the “matrix factorization (MF) + factorization machine (FM)” framework for the latter. For the MF part, we obtain the user-item similarity matrix from each metagraph and then apply low-rank matrix approximation to obtain latent features for both users and items. For the FM part, we propose to apply FM with Group lasso (FMG) on the features obtained from the MF part to train the recommendation model and at the same time identify the useful metagraphs. Experimental results from two large real-world datasets, i.e., Amazon and Yelp, show that our proposed approach is better than FM and other state-of-the-art HIN-based recommendation methods.

Finally, besides metagraph, we further propose the Motif Enhanced MetaPath (MEMP) method for computing the similarities between users and items in HIN-based RSs. Motif is a local structure that can capture higher-order relations among nodes in homogeneous graphs. We argue that such higher-order relations also exist among nodes of same types in HIN. Thus, existing metapath based similarities can also be enhanced by integrating these motif-based higher-order relations. After computing the MEMP based similarities between users and items, we apply our proposed “MF+FM” framework to fuse the similarities for rating prediction. Experiments have been conducted on two real-word datasets, Epinions and CiaoDVD, and the results demonstrate the effectiveness of MEMP in HIN-based RSs.