**RecDCL: Dual Contrastive Learning for Recommendation** **Authors:** Dan Zhang **Keywords:** Recommender Systems, Self-supervised Learning, Batch-wise Contrastive Learning, Feature-wise Contrastive Learning **Abstract:** Self-supervised learning (SSL) has achieved significant success in mining user-item intentions for collaborative filtering. Contrastive learning (CL) based SSL addresses data sparsity by contrasting embeddings between raw and augmented data. However, existing CL-based methods primarily focus on batch-wise contrastive learning, failing to leverage feature dimension regularity. This paper investigates the combination of batch-wise CL (BCL) and feature-wise CL (FCL) for recommendation. Theoretical analysis reveals that combining BCL and FCL helps eliminate redundant solutions without missing optimal solutions. We propose RecDCL, a dual contrastive learning framework. The FCL objective optimizes user-item positive pairs and uniform distributions within users and items using a polynomial kernel. The BCL objective enhances representation robustness by generating contrastive embeddings on output vectors. Extensive experiments on four benchmarks and one industry dataset demonstrate that RecDCL outperforms state-of-the-art GNNs-based and SSL-based models, achieving improvements of up to 5.65% in Recall@20. **Contributions:** - Theoretical analysis reveals the connection between BCL and FCL and demonstrates their cooperative benefits. - RecDCL is proposed, integrating FCL and BCL objectives to learn informative representations. - Extensive experiments validate the effectiveness of RecDCL, showing significant performance improvements over state-of-the-art models. **Introduction:** - BCL and FCL are two major types of contrastive learning objectives. - BCL focuses on maximizing similarity between positive pairs and minimizing similarity between negative pairs. - FCL emphasizes decorrelating embedding components in the feature-wise dimension. - Previous works have explored the connection between BCL and FCL but lacked a native interpretation. - This paper reveals the inherent connection between BCL and FCL and demonstrates their cooperative benefits. **Methodology:** - RecDCL combines FCL and BCL objectives to enhance representation learning. - FCL objective (UIBT and UUII) captures alignment and uniformity in user-item interactions. - BCL objective (Basic BCL and Advanced BCL) enhances representation robustness through data augmentation. **Experiments:** - RecDCL outperforms state-of-the-art models on four public datasets and an industrial dataset. - Ablation studies validate the effectiveness of each component in RecDCL. - Industrial results show significant improvements in Recall@20 and NDCG@20. **Conclusion:** RecDCL effectively combines BCL and FCL to learn informative representations for recommendation, demonstrating superior performance compared to state-of-the-art models.**RecDCL: Dual Contrastive Learning for Recommendation** **Authors:** Dan Zhang **Keywords:** Recommender Systems, Self-supervised Learning, Batch-wise Contrastive Learning, Feature-wise Contrastive Learning **Abstract:** Self-supervised learning (SSL) has achieved significant success in mining user-item intentions for collaborative filtering. Contrastive learning (CL) based SSL addresses data sparsity by contrasting embeddings between raw and augmented data. However, existing CL-based methods primarily focus on batch-wise contrastive learning, failing to leverage feature dimension regularity. This paper investigates the combination of batch-wise CL (BCL) and feature-wise CL (FCL) for recommendation. Theoretical analysis reveals that combining BCL and FCL helps eliminate redundant solutions without missing optimal solutions. We propose RecDCL, a dual contrastive learning framework. The FCL objective optimizes user-item positive pairs and uniform distributions within users and items using a polynomial kernel. The BCL objective enhances representation robustness by generating contrastive embeddings on output vectors. Extensive experiments on four benchmarks and one industry dataset demonstrate that RecDCL outperforms state-of-the-art GNNs-based and SSL-based models, achieving improvements of up to 5.65% in Recall@20. **Contributions:** - Theoretical analysis reveals the connection between BCL and FCL and demonstrates their cooperative benefits. - RecDCL is proposed, integrating FCL and BCL objectives to learn informative representations. - Extensive experiments validate the effectiveness of RecDCL, showing significant performance improvements over state-of-the-art models. **Introduction:** - BCL and FCL are two major types of contrastive learning objectives. - BCL focuses on maximizing similarity between positive pairs and minimizing similarity between negative pairs. - FCL emphasizes decorrelating embedding components in the feature-wise dimension. - Previous works have explored the connection between BCL and FCL but lacked a native interpretation. - This paper reveals the inherent connection between BCL and FCL and demonstrates their cooperative benefits. **Methodology:** - RecDCL combines FCL and BCL objectives to enhance representation learning. - FCL objective (UIBT and UUII) captures alignment and uniformity in user-item interactions. - BCL objective (Basic BCL and Advanced BCL) enhances representation robustness through data augmentation. **Experiments:** - RecDCL outperforms state-of-the-art models on four public datasets and an industrial dataset. - Ablation studies validate the effectiveness of each component in RecDCL. - Industrial results show significant improvements in Recall@20 and NDCG@20. **Conclusion:** RecDCL effectively combines BCL and FCL to learn informative representations for recommendation, demonstrating superior performance compared to state-of-the-art models.