RecDCL: Dual Contrastive Learning for Recommendation This paper proposes RecDCL, a dual contrastive learning framework for recommendation systems that combines batch-wise contrastive learning (BCL) and feature-wise contrastive learning (FCL). The framework aims to eliminate redundant solutions in representation learning while preserving optimal solutions. The BCL objective generates contrastive embeddings to enhance representation robustness, while the FCL objective optimizes uniform distributions within users and items using a polynomial kernel to drive representations to be orthogonal. Extensive experiments on four public datasets and one industry dataset show that RecDCL outperforms state-of-the-art GNN-based and SSL-based models, achieving up to 5.65% improvement in Recall@20. The source code is publicly available. The paper analyzes the relationship between BCL and FCL, revealing that combining them helps eliminate redundant solutions without missing optimal solutions. It also explores the effectiveness of FCL in self-supervised learning, showing that it can improve recommendation performance by promoting feature-wise uniformity. The paper further investigates the benefits of combining BCL and FCL, demonstrating that they complement each other in enhancing representation learning. The proposed RecDCL framework includes two main objectives: a RecFCL objective for driving representations to be orthogonal and a Rec-BCL objective for enhancing representation robustness. The framework uses a polynomial kernel to optimize feature-wise uniformity and employs historical embeddings to generate contrastive views for representation learning. The framework is evaluated on four public datasets and one industry dataset, showing significant improvements in recommendation performance. The paper concludes that combining BCL and FCL provides a more effective approach to self-supervised learning for recommendation systems. It also highlights the importance of feature-wise objectives in improving recommendation performance and suggests that future research should explore other CL-based training objectives that favor feature-wise perspectives.RecDCL: Dual Contrastive Learning for Recommendation This paper proposes RecDCL, a dual contrastive learning framework for recommendation systems that combines batch-wise contrastive learning (BCL) and feature-wise contrastive learning (FCL). The framework aims to eliminate redundant solutions in representation learning while preserving optimal solutions. The BCL objective generates contrastive embeddings to enhance representation robustness, while the FCL objective optimizes uniform distributions within users and items using a polynomial kernel to drive representations to be orthogonal. Extensive experiments on four public datasets and one industry dataset show that RecDCL outperforms state-of-the-art GNN-based and SSL-based models, achieving up to 5.65% improvement in Recall@20. The source code is publicly available. The paper analyzes the relationship between BCL and FCL, revealing that combining them helps eliminate redundant solutions without missing optimal solutions. It also explores the effectiveness of FCL in self-supervised learning, showing that it can improve recommendation performance by promoting feature-wise uniformity. The paper further investigates the benefits of combining BCL and FCL, demonstrating that they complement each other in enhancing representation learning. The proposed RecDCL framework includes two main objectives: a RecFCL objective for driving representations to be orthogonal and a Rec-BCL objective for enhancing representation robustness. The framework uses a polynomial kernel to optimize feature-wise uniformity and employs historical embeddings to generate contrastive views for representation learning. The framework is evaluated on four public datasets and one industry dataset, showing significant improvements in recommendation performance. The paper concludes that combining BCL and FCL provides a more effective approach to self-supervised learning for recommendation systems. It also highlights the importance of feature-wise objectives in improving recommendation performance and suggests that future research should explore other CL-based training objectives that favor feature-wise perspectives.