MDGCL: Message Dropout Graph Contrastive Learning for Recommendation

In the realm of recommendation systems, the representation learning of users and items plays a pivotal role. The advent of Graph Neural Networks (GNN) has propelled Graph Collaborative Filtering (GCF) to new heights by effectively capturing high-order connectivities. However, the persistent challenge of matrix sparsity in collaborative filtering continues to impede optimal performance enhancement. While Graph Contrastive Learning (GCL) has emerged as a solution to mitigate data sparsity by extracting general features from raw data, its conventional methods, involving node and edge dropout for graph augmentation, often result in the loss of crucial graph information, thereby diminishing model performance. Addressing this, we introduce a novel technique termed MessageDropout, implemented during the message passing phase. This approach offers a more refined granularity and ensures the retention of key information within the graph. Additionally, our Graph Contrastive Learning (GCL) method uniquely eliminates the need for pre-graph augmentation and avoids the requirement of two additional forward trainings in each mini-batch. We name our proposed method MDGCL. Experiments on three public datasets have proven the effectiveness of our proposed model. The code is released at https://github.com/anorepo/MDGCL .