Invariant Graph Contrastive Learning for Mitigating Neighborhood Bias in Graph Neural Network Based Recommender Systems

Graph neural network(GNN) is playing a vital role in recommender systems to capture complex dynamics of user-item interactions through user-item bipartite graphs. Despite the success of existing GNN-based recommender systems, they generally suffer from the neighborhood bias problem, which breaks the homophily assumption in various real-world recommendation scenarios. The neighborhood bias stems from the mixed complex local patterns caused by the diverse user preferences and the multifaceted item attributes, where two users with distinct preferences might be wrongly connected via a shared item node. To this end, we introduce the Invariant Graph Contrastive Learning (IGCL) framework, which is the first work to address the neighborhood bias from the perspective of invariant learning. We discern the local pattern distributions based on the contrastive user representations, and employ invariant learning to alleviate the neighborhood bias by eliminating the shared variant vectors. Our experiments on three real-world public datasets demonstrate that the IGCL framework significantly outperforms existing baselines, offering a promising solution to overcome the neighborhood bias in GNN-based recommender systems. The source code of our work is available at https://github.com/ShallowDream-97/IGCL .