Graph Learning in Recommender Systems: Toward Structures and Causality

Recently, researchers have been exploring the application of machine learning techniques to learn the characteristics of graphs from raw data, a field known as graph learning. Graph learning is particularly useful in Recommender Systems (RS), where objects such as users, items, and their attributes are closely connected and influence each other through various relationships. These systems are highly effective in mitigating the problem of information overload on the internet, by enabling users to find relevant products, songs, movies, or news quickly, based on their preferences and without extensive searching. The basic idea behind an RS is to model the relationships between users and items based on their past interactions. With the advancement of graph learning techniques, exploring and utilizing homogeneous and heterogeneous relationships within graphs holds immense potential for enhancing the effectiveness of RS. In this dissertation, I introduce four recent works about graph learning for recommender systems. In the first work, I propose a model named Dynamic Graph Collaborative Filtering (DGCF) based on dynamic graph learning to learn the changes in user and item embeddings. In the second work, a GNN-based model named RAM-GNN is proposed to learn the complex relations in the user and item graphs, respectively. In the third graph, I propose a causal graph indicating the frequency bias in the next-basket recommendation and a model based on machine learning is derived to solve the problem. In the last work, I propose a model based on the graph attention network to learn the item graph in session-based recommendation in a hyperbolic space.