Adaptive Popularity Debiasing Aggregator for Graph Collaborative Filtering

The graph neural network-based collaborative filtering (CF) models user-item interactions as a bipartite graph and performs iterative aggregation to enhance performance. Unfortunately, the aggregation process may amplify the popularity bias, which impedes user engagement with niche (unpopular) items. While some efforts have studied the popularity bias in CF, they often focus on modifying loss functions, which can not fully address the popularity bias in GNN-based CF models. This is because the debiasing loss can be falsely backpropagated to non-target nodes during the backward pass of the aggregation.

In this work, we study whether we can fundamentally neutralize the popularity bias in the aggregation process of GNN-based CF models. This is challenging because 1) estimating the effect of popularity is difficult due to the varied popularity caused by the aggregation from high-order neighbors, and 2) it is hard to train learnable popularity debiasing aggregation functions because of data sparsity. To this end, we theoretically analyze the cause of popularity bias and propose a quantitative metric, named inverse popularity score, to measure the effect of popularity in the representation space. Based on it, a novel graph aggregator named APDA is proposed to learn per-edge weight to neutralize popularity bias in aggregation. We further strengthen the debiasing effect with a weight scaling mechanism and residual connections. We apply APDA to two backbones and conduct extensive experiments on three real-world datasets. The results show that APDA significantly outperforms the state-of-the-art baselines in terms of recommendation performance and popularity debiasing.