Data Sparsity in Recommender Systems: Collaborative Neighbors Enrichment

In recommender systems, collaborative filtering is the most predominant idea for modeling users' preferences on items. Collaborative filtering learns user-user similarities and item-item correlations by representing users/items with their interacted items/users (i.e., collaborative neighbors). However, the data sparsity characteristic of user-item interactions poses a great challenge to users and items modeling. Moreover, most users and items have limited interaction, which means users and items both follow the power-law distribution based on the number of interactions. This data sparsity issue limits the representation learning of users/items due to insufficient collaborative neighbors. In this dissertation, I will introduce our recent work on empowering different methods in alleviating the data sparsity problem in recommender systems, in perspectives of data augmentations, novel metric learning, and the usage of additional information to help expand the collaborative neighbors. In this first part, I present an approach to augment pseudo-history for short user sequences reversely. In the second and third parts, a novel metric based on Wasserstein distance is introduced to enhance neighbors via collaborative transitivity. In the fourth part, a multi-relational Transformer is proposed to incorporate additional item relationships. Lastly, I will introduce to apply Wasserstein distance into contrastive learning to enhance the robustness and efficiency of data augmentations.