Graph neural networks for improvement recommender system

Abstract

The field of recommender systems is a significant branch of data mining, with extensive applications in both research and practical contexts. These systems leverage user and item interaction data to generate personalized item recommendations, enhancing user experiences in e-commerce by promoting items that align with user preferences. Beyond e-commerce, recommender systems are also applied in recommending books, movies, courses, and websites.

User recommendations are primarily determined by the similarity between users, assessed using techniques such as collaborative filtering (CF), matrix factorization (MF), and singular value decomposition (SVD). These techniques typically rely on matrices of past user ratings. In graph data mining, interactions between users and items can be represented as a graph, enabling more extensive and effective interaction filtering techniques. Graph Neural Network (GNN) models are crucial for representing and extracting insights from graph data.

Social relationships significantly influence shopping habits, with recommendations from friends often having a greater impact than those from other sources. Integrating social connections with user similarity, based on shared shopping patterns, presents challenges due to the noise and variability in different datasets. This complexity is further amplified when considering relationships between items, such as geographic location, product type, buyer comments, and visit time sessions.

Summarized, the integration of advanced techniques such as GNNs and the consideration of social connections and item relations offer promising advancements in the effectiveness of recommender systems, despite the inherent challenges.