AdaDCL: An Adaptive Disentangled Contrastive Recommendation Method

Graph contrastive learning has demonstrated outstanding performance in addressing the issue of label scarcity. It still has two limitations: 1) the stacking of graph layers often leads to over-smoothing, hard to distinguish the embeddings of distinct nodes; and 2) traditional algorithms typically model user preferences with a unified intention, neglecting the multifaceted and fine-grained motivations behind user-item interactions. To overcome these shortcomings, we propose an adaptive disentangled contrastive learning (AdaDCL) method tailored for recommendation systems. First, we perform disentangled modeling of global information intentions and introduce a cross-view contrastive learning task, employing a parameterized mask generator for adaptive augmentation. Second, we employ a layer attention mechanism to counter over-smoothing in GNNs, ensuring that meaningful semantic features are preserved across layers. Third, an adaptive hardness negative sampling (AHNS) strategy dynamically selects negative samples based on their hardness levels, reducing the risk of false positives and negatives while enhancing the robustness of contrastive learning. Comprehensive experiments on three benchmark datasets, including Gowalla, and comparisons against twelve state-of-the-art baseline models (e.g., DisenHAN), demonstrate that AdaDCL surpasses the classic LightGCN by 5.89% in Recall@20 and over 7% in NDCG@20 on the Gowalla dataset. These results highlight the effectiveness and generalizability of our approach.