Single-Target and Dual-Target Cross-Domain Recommendation

To address the data sparsity problem in recommender systems, cross-domain recommendation (CDR) has in recent years leveraged the relatively richer information from a richer (source) domain to only improve the recommendation performance in a sparser (target) domain with sparser information. Existing CDR approaches either directly replace a part of the latent representation of users/items in the sparser domain with the corresponding latent representation in the richer domain, or they map the latent representation of common users/items in the richer domain to fit those in the sparser domain.

First, finding an accurate mapping of the latent factors across domains is crucial for enhancing recommendation accuracy for CDR. However, this is a challenging task because of the complex relationships that exist between the latent factors of the source and the target domains or systems. To this end, this thesis proposes a deep framework for both cross-domain and cross-system recommendations (DCDCSR) based on matrix factorisation (MF) models and a fully connected deep neural network (DNN). Specifically, DCDCSR first employs the MF models to generate user and item latent factors and then employs the DNN to map the latent factors across domains or systems. More importantly, this approach considers the rating sparsity degrees of individual users and items in different domains or systems and uses them to guide the DNN training process for utilising the rating data more effectively.

Second, the existing CDR approaches are single-target approaches. However, each of the two domains may be relatively richer in certain types of information (e.g., ratings, reviews, user profiles, item details and tags). If such information can be leveraged well, it is thus possible to simultaneously improve the recommendation performance in both domains (i.e., dual-target CDR) rather than in a single-target domain only. Thus, to achieve dual-target CDR, this thesis proposes a new framework for dual-target cross-domain recommendation (DTCDR). In the DTCDR framework, rating and multi-source content information are first extensively used to generate rating and document embeddings of users and items. Then, based on multi-task learning (MTL), an adaptable embedding-sharing strategy is designed to combine and share the embeddings of common users across domains, with which DTCDR can improve the recommendation performance on both richer and sparser (i.e., dual-target) domains simultaneously.

Third, inspired by DTCDR, this thesis attempts to further improve the recommendation performance in both domains. There are two new challenges: (1) how to generate more representative user and item embeddings, and (2) how to effectively optimise the user/item embeddings in each domain. To address these challenges, this thesis proposes a graphical and attentional framework, called GA-DTCDR. In the GA-DTCDR framework, two separate heterogeneous graphs are first constructed based on the rating and content information from the two domains to generate more representative user and item embeddings. Then, an element-wise attention mechanism is proposed for effectively combining the embeddings of common users learned from both domains. Both steps significantly enhance the quality of user and item embeddings and thus improve the recommendation accuracy in each domain.

All the above approaches proposed in this thesis have been validated and evaluated by theoretical analysis and extensive experiments conducted on real-world datasets. The experimental results demonstrate that the proposed methods significantly improve the recommendation accuracies in both the richer and sparser domains, and that these approaches outperform the state-of-the-art single-domain recommendation approaches, single-target CDR approaches, and dual-target CDR approaches in terms of recommendation accuracy.