DCL4Rec: An Effective Debiased Contrastive Learning Framework for Long-Tail Recommendation

Sequential recommendation offers a promising avenue for long-tail scenarios by modeling user behavior sequences, enabling the discovery of interest shifts and reducing the influence of popularity bias. This helps improve the fairness and accuracy of recommendations. However, the inherent imbalance of long-tail item distributions and the sparsity of user interactions often lead to representation bias and overfitting, ultimately degrading performance. To address these challenges, we propose DCL4Rec, a debiased contrastive learning framework tailored for long-tail recommendation. DCL4Rec leverages the SASRec model as a sequential encoder to capture user behavior patterns and generate both positive and in-batch negative samples. To further enhance robustness, we introduce Gaussian negative sampling, which mitigates the effects of noisy data and promotes uniformity in the representation space. Central to our framework is a debiasing mechanism integrated into the contrastive learning objective, combining contrastive filtering with representation regularization. This strategy effectively counters the popularity bias and enables more balanced representation learning across both head and tail items. Extensive experiments on real-world datasets demonstrate that DCL4Rec significantly outperforms existing state-of-the-art methods in long-tail recommendation tasks. The results confirm the framework's effectiveness in addressing key limitations of conventional contrastive and sequential recommendation methods, highlighting its potential to deliver fairer, more diverse, and higher-quality recommendations in practice.