Explainable Recommendation in Graph Neural Networks Using Propagation Path Analysis and Counterfactual Graph Editing

Abstract

Graph Neural Networks (GNNs), particularly LightGCN, have achieved strong performance in collaborative filtering-based recommendation systems. However, their black-box nature makes it difficult to explain why specific items are recommended, limiting trust and adoption in user-facing applications. In this paper, we propose a post-hoc explainability framework for LightGCN combining two complementary techniques: Propagation Path Analysis, which decomposes recommendation scores by propagation layer to attribute influence to specific training interactions, and Counterfactual Graph Editing, which identifies the most influential user-item edges through structural sensitivity analysis and targeted edge removal. We evaluate on MovieLens-1M against three baselines including random edge removal, degree-based attribution, and LIME. Our method achieves a mean absolute score drop of 0.0322, representing a 3.02× lift over random, a 1.26× lift over LIME, and a 5.35× lift over degree-based attribution. Faithfulness evaluation against post-fine-tuning ground truth yields Pearson r = 0.536, confirming that structural sensitivity reliably identifies influential edges without model retraining. Layer contribution analysis reveals that LightGCN recommendations are predominantly driven by direct interactions (Layer 0) and 1-hop neighbours (Layer 1), with deeper layers contributing progressively less.