Recommendation System Debiasing Algorithm Based on Causal Inference and Graph Neural Network
Year: 2026
Type: conference-paper
Abstract: To address the pervasive selection bias and exposure bias in recommender systems, this paper proposes a causal de-biased graph neural network model termed CI-GNN, which integrates causal inference with graph-based recommendation. Specifically, a structural causal model is first constructed over user, item, exposure, and popularity variables, and the observed interaction data are de-biased through backdoor adjustment and counterfactual reasoning. The estimated causal weights are then embedded into the graph propagation process of LightGCN, enabling the diffusion of causal signals over the user–item interaction graph and yielding more accurate representations of users’ true preferences. Extensive simulated experiments on MovieLens-25M, Amazon, and Last.fm datasets demonstrate that CI-GNN consistently outperforms MF, NeuMF, LightGCN, and CausalRec in terms of Recall@20 and NDCG@20. Moreover, significant improvements are observed in long-tail recommendation scenarios, verifying the effectiveness of causal modeling in mitigating popularity-induced bias and enhancing recommendation fairness.
Authors: Shaoxi Li, Jiaoyan Chen, Jianjia Zhang
Institution: Guangdong University of Science and Technology
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204989286