Exploiting side information for effective recommendation

• Main Author: Sha, Xiao
• Other Authors: Zhang Jie
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/155781
• Institution: Nanyang Technological University

Recommender systems have become an essential tool in online applications to help alleviate the information overload problem. Much research effort has been devoted to traditional collaborative filtering (CF), which, however, suffers from the data sparsity and cold start problems. To ease these issues, the side information, such as social networks and item attributes, has been incorporated into CF for improving the recommendation performance. However, it is still challenging to effectively extract and represent the auxiliary information due to the complex network structure and data heterogeneity. Therefore, in this dissertation, we focus on developing delicate algorithms to fully exploit the side information for effective recommendation. Social networks, which provide the homogeneous side information of users, have proven to be effective for high-quality item recommendation. In the social network, users may establish a variety of social connections based on different motivations, such as friends sharing similar interests or activities, work-related friends, school friends and families. Hence, we argue that social relations may have multiple latent facets, and different facets may have varying contributions to item recommendation. Most social recommenders, however, largely neglect the multi-facet social relations latent in the social network, which is insufficient to capture the user preferences over items. We, therefore, propose a disentangled social recommendation (DSR) framework to exploit the multi-facet social relations for enhanced item recommendation. Specifically, DSR explicitly disentangles the social relations into multiple facets, and encodes the social influence under each facet into disentangled user embeddings in the social network. The multiple user embeddings are then aggregated via a facet-level attention mechanism, which distinguishes the effective facets for better inferring user interests over items. Empirical study shows the superiority of DSR against state-of-the-art methods and its potential in easing the data sparsity issue. In addition to homogeneous social networks, the knowledge graph (KG) has attracted increasing attention in recommendation, as it provides abundant side information of heterogeneous entities and relations. By exploring the interlinks of KGs, the connectivities between users and items help reveal their underlying relationship, which are complementary to the user--item interactions. To fully exploit the heterogeneous information in KGs, we thus design a novel hierarchical attentive knowledge graph embedding (HAKG) framework for enhanced recommendation. Specifically, HAKG first extracts the expressive subgraphs that link user--item pairs to characterize their connectivities, which accommodate both the semantics and topology of KGs. The subgraphs are then encoded via a hierarchical attentive subgraph encoding to generate effective subgraph embeddings for enhanced user preference prediction. Extensive experiments demonstrate the effectiveness of HAKG over the state-of-the-art methods in terms of recommendation performance, especially for inactive users with sparse interactions over items. In recent years, group buying (GB), as a new business paradigm in social e-commerce, has sprung up and turned out to be an immediate success. In this business model, a user can launch a GB as an initiator to share her interested product with social friends. The GB is clinched once enough friends join in as participants to co-purchase the shared product. As such, there are heterogeneous relations in the GB scenario, e.g., social relations, initiator/participant-product interactions, which may have varying contributions to the user preference inference. To fully exploit the heterogeneous relations in GB, we thus propose a joint product-participant recommendation (J2PRec) framework for enhanced GB recommendation. Specifically, J2PRec first designs a relational graph embedding module, which effectively encodes the heterogeneous relations in GB for learning enhanced user and product embeddings. It then jointly learns the product and participant recommendation tasks under a probabilistic framework to maximize the GB likelihood, i.e., boost the success rate of a GB. Experiments on real-world datasets demonstrate the superiority of J2PRec for GB recommendation. To sum up, in this dissertation, we propose a series of novel recommendation approaches by exploiting the various side information from homogeneous to heterogeneous for high-quality recommendation. Extensive experiments on real-world datasets show the effectiveness of our proposed frameworks in recommendation performance.