融合对抗混合负采样的图对比学习推荐算法研究OA

Contrastive learning has shown effectiveness in alleviating data sparsity and improving representation learning in recommender systems.However,existing methods mainly focus on integrating contrastive learning with specific recom-mendation tasks,while overlooking the impact of negative sample quality and embedding robustness.To address this issue,a graph contrastive learning recommendation model with adversarial mixed negative sampling,termed AMixGCL,is proposed.The model constructs mixed adversarial negative samples in the embedding space to provide more informative and challenging training signals,thereby enhancing discriminative capability.In addition,contrastive learning is conducted across multi-level graph embeddings to fully exploit structural information,which effectively mitigates the over-smoothing problem and improves representation diversity.An adaptive noise-based augmentation strategy is further employed to generate consistent contrastive views,and the recommendation objective is jointly optimized with the contrastive learning task.Comprehensive experiments are conducted on three large-scale and highly sparse benchmark datasets,namely Yelp2018,Amazon-kindle,and Alibaba-iFashion.The experimental results demonstrate that the proposed method achieves relative improvements of 24%,22%,and 38%in terms of Recall@20,and 25%,26%,and 42%in terms of NDCG@20,respectively,compared with the baseline model.These results indicate that the proposed approach consistently outperforms existing recommendation methods,validating its effectiveness and superiority in recommendation tasks.