Causality-aware Graph Aggregation Weight Estimator for Popularity Debiasing in Top-K Recommendation

TL;DR

The paper tackles popularity bias in graph-based top-K recommendations by reframing graph aggregation as a backdoor adjustment problem in causal inference. It introduces CAGED, an encoder–decoder model that learns unbiased history likelihoods to adjust aggregation weights via an ELBO objective, and integrates a momentum-based update to balance debiasing with representation learning. Empirical results on three real datasets show that CAGED improves both overall recommendation performance and niche-item coverage while maintaining robust performance on popular items. The work advances graph-based debiasing by providing causal grounding, a learnable weighting mechanism, and a training strategy that stabilizes optimization, with practical impact for fairer, more diverse recommendations.

Abstract

Graph-based recommender systems leverage neighborhood aggregation to generate node representations, which is highly sensitive to popularity bias, resulting in an echo effect during information propagation. Existing graph-based debiasing solutions refine the aggregation process with attempts such as edge reconstruction or weight adjustment. However, these methods remain inadequate in fully alleviating popularity bias. Specifically, this is because 1) they provide no insights into graph aggregation rationality, thus lacking an optimality guarantee; 2) they fail to well balance the training and debiasing process, which undermines the effectiveness. In this paper, we propose a novel approach to mitigate popularity bias through rational modeling of the graph aggregation process. We reveal that graph aggregation is a special form of backdoor adjustment in causal inference, where the aggregation weight corresponds to the historical interaction likelihood distribution. Based on this insight, we devise an encoder-decoder architecture, namely Causality-aware Graph Aggregation Weight Estimator for Debiasing (CAGED), to approximate the unbiased aggregation weight by optimizing the evidence lower bound of the interaction likelihood. In order to enhance the debiasing effectiveness during early training stages, we further design a momentum update strategy that incrementally refines the aggregation weight matrix. Extensive experiments on three datasets demonstrate that CAGED outperforms existing graph-based debiasing methods. Our implementation is available at https://github.com/QueYork/CAGED.

Figures (6)

• Figure 1: A toy example of popularity bias. Nodes $u_1$, $i_1$ are popular while the rest are niche. In the aggregation process of this bipartite graph (left), popular nodes aggregate the entire opposite set while niche nodes can only incorporate their common popular neighbor (right).
• Figure 2: (a): A typical causal graph example. (b): Causal graph of general recommendation scenario.
• Figure 3: An illustration of CAGED architecture. Embeddings of the convolution center $u$ and one of its neighbors $x$ are forwarded into the encoder and decoder modules respectively. The ELBO output from two modules forms the CAGED-generated weight $\mathbf{W_{CAGED}[u, x]}$ to incrementally update the original aggregation weight $\mathbf{A[u, x]}$.
• Figure 4: Visualization of inverse item popularity before, during, and after CAGED update.
• Figure 5: Recommendation performances of CAGED under different $\boldsymbol{\lambda}$s.