نوع مقاله : مقاله پژوهشی
نویسندگان
1 هیئت علمی گروه علوم کامپیوتر، دانشگاه سیستان و بلوچستان
2 گروه علوم کامپیوتر، دانشگاه تهران
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Leveraging social networks as an auxiliary information source is considered an effective solution to overcome the data sparsity challenge in recommender systems. However, real-world social relations are often noisy, and the assumption of "perfect preference alignment among friends" does not always hold true. Recently, Graph Contrastive Learning (GCL) has garnered attention to tackle this noise; nevertheless, conventional approaches relying on stochastic methods such as edge dropping distort the graph's semantic structure and yield unstable representations. To address these limitations, this paper proposes a novel framework based on graph contrastive learning utilizing adversarial noise. In this architecture, instead of perturbing the graph structure, targeted perturbations are injected into the continuous representation space using the Fast Gradient Method (FGM) to generate "hard" and robust contrastive views. Furthermore, a dual contrastive learning approach is developed, comprising intra-domain alignment (to extract the pure features of each graph) and cross-domain alignment (to purposively transfer useful social signals to the interaction domain). Empirical evaluations on four real-world datasets (Yelp, Epinions, Trustfilm, Ciao) demonstrate that the proposed model outperforms state-of-the-art baseline methods, achieving an average performance improvement of $8.5\%$ in the Recall metric and $7.2\%$ in the NDCG metric. These results verify the effectiveness of employing adversarial noise in extracting robust representations and improving recommendation accuracy.
کلیدواژهها [English]
)