<p>Contrastive learning has gained significant attention in the field of recommender systems due to its ability to learn highly expressive representations with limited labels. However, historical user–item interaction data used for recommender systems often contain confounders, thereby establishing spurious correlations between user preferences and confounders during self-supervised training and misleading recommender systems to use these correlations as shortcuts for generating recommendations. Existing approaches for debiasing usually involve manually identifying observed confounders, but they are often tailored to specific situations and overlook latent confounders. To address this challenging problem, we propose a Deconfounding Graph Contrastive Learning (DeGCL) method to provide deconfounding recommendations by adjusting for a learned deconfounding representation from interaction data, using the back-door adjustment strategy. DeGCL learns the representation to capture latent confounding effects in observational data between users and items. It artificially adds interactions and noise to create contrastive views, which help deconfound the model. By adjusting for the learned representation, DeGCL mitigates latent confounding effects in training downstream recommendation models. Experiments on two real-world datasets demonstrate that our method outperforms state-of-the-art methods, suggesting its potential to provide more effective recommendations in practice.</p>