Diffusion-Enhanced Identifiable Confounder Modeling for Debiased Recommendation

In real-world recommender systems, exposure bias, data sparsity, and latent confounding factors can substantially distort the authenticity of user–item interactions, leading to predictions that deviate from users’ true preferences. To address these challenges, this paper proposes a unified debiased recommendation framework, DiffCL, which integrates Matrix Factorization (MF), Identifiable Variational Autoencoders (iVAE), Diffusion Models, and Contrastive Learning. Specifically, MF is first employed to capture explicit collaborative signals, while iVAE is introduced to construct identifiable prior representations that encode users’ causal preferences. Subsequently, a diffusion-based denoising process is applied to the latent variables for fine-grained augmentation, enabling the model to obtain robust and high-quality hidden representations through multi-step generation and inversion. In addition, contrastive learning is incorporated as a regularization mechanism to enforce semantic consistency between user and item embeddings, further enhancing the discriminability and structural organization of the latent space. Extensive experimental results on multiple real-world datasets demonstrate that the proposed model consistently outperforms state-of-the-art baselines, particularly under highly biased and sparse settings, exhibiting stronger generalization capability and training stability. This study provides new insights into the integration of causal representation learning and diffusion-based generative modeling in recommender systems, and lays a solid foundation for building more reliable, fair, and robust recommendation models.