Integrating a Sequential Model Into GNN-Based Social Recommendation for Relieving Over-Smoothing Problem

In recent years, social recommendation systems have increasingly integrated Graph Neural Networks (GNNs) to capture social relationships among users from both direct and higher-order neighbors. However, these state-of-the-art models face limitations regarding the number of iterations they can propagate across the graph. Messages from higher-order neighbors often dilute the unique characteristics of the target node. This causes node representations to become similar and indistinguishable, leading to a phenomenon known as an over-smoothing problem. Several works have attempted to address this problem with the expectation of preserving user characteristics. They add auxiliary information to the GNN-based social recommendation models without considering the dynamic preferences of users, which frequently change over time. Therefore, those works still struggle with the same problem after propagating in a few iterations. To mitigate this phenomenon, we propose a novel approach to integrate user characteristics extracted from users’ historical interactions using a sequential model to slow down the node convergence. Integrating user characteristics into the GNN-based social recommendations will allow the model to capture user preferences that dynamically change over time. Consequently, the proposed model will be able to collect informative data from higher-order neighbors over social networks for user representation learning, leading to better performance and effectiveness of the recommendations. Our experimental results demonstrate that the proposed model outperforms state-of-the-art social recommendation models across three benchmark datasets.