| 摘 要: 针对现有图卷积推荐系统在捕捉用户个性化项目特征敏感性方面的不足,提出一种联合多粒度流行度感知的图协同过滤推荐模型。首先,采用多粒度流行度特征建模,有效缓解用户与项目交互矩阵的高稀疏性问题,从而揭示用户偏好在不同流行度特征上的细微差异。其次,引入对比学习辅助任务,通过自监督学习增强节点表达。最后,联合优化有监督推荐任务与自监督学习辅助任务,获得更加准确的推荐。在Gowalla、Yelp2018和Amazon-Book3个公开数据集上进行了对比实验。实验结果表明,该模型在 Recal@20 指标上分别取得了1.64%、3.54%、4.38%的提升,在 NDCG@20 指标上分别取得了 2.80%、6.69%、8.42%的提升,充分证明了模型的有效性和优越性。 |
| 关键词: 图卷积网络 自监督学习 协同过滤 多粒度流行度 对比学习 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(61977039) |
|
| Joint Multi-grained Popularity-Aware Graph Collaborative Filtering Recommendation Model |
|
XU Longhai, GU Yiran
|
(College of Automation & College of Artificial Intelligence, Nanjing University of Posts and Telecommunications, Nanjing 210023, China)
1222056229@njupt.edu.cn; guyr@njupt.edu.cn
|
| Abstract: Addressing the limitations of existing graph convolutional recommendation systems in capturing users’personalized sensitivity to item features, this paper proposes a joint mult-i grained popularity-aware graph collaborative filtering recommendation model. Firstly, the model employs mult-i grained popularity feature modeling to effectively alleviate the high sparsity of use-r item interaction matrices, thereby revealing nuanced variations in user preferences
across different popularity characteristics. Secondly, a contrastive learning auxiliary task is introduced to enhance node representations through sel-f supervised learning. Finally, joint optimization of the supervised recommendation task and the sel-f supervised auxiliary task yields more accurate recommendations. Comparative experiments on three public datasets (Gowalla, Yelp2018, and Amazon-Book) demonstrate performance improvements: Recall @ 20 increased by 1.64%, 3.54% and superiority, and 4.38%, while NDCG@20 rose by 2.80%, 6.69%, and 8.42% respectively, validating the model’s effectiveness. |
| Keywords: graph convolutional networks sel-f supervised learning collaborative filtering mult-i grained popularity contrastive learning |
