融合原型对比与特征筛选的图协同过滤模型

A graph collaborative filtering model combining prototype comparison and feature filtering

图卷积在协同过滤推荐系统上取得了巨大的成功,但在真实的推荐场景中基于协同过滤的推荐方法往往会受到稀疏数据的影响,同时现有的图协同过滤方法又普遍存在对用户-项目交互信息的分析和利用不彻底的问题,如未对交互特征中的噪声进行处理,这些问题使得现有方法的推荐效果不理想。针对上述问题,提出了一种融合原型对比与特征筛选的图协同过滤模型,在对交互信息进行噪声特征过滤的同时,利用提出的原型对比学习任务捕捉节点间的潜在联系,以此增强用户和项目的表示。在3个真实的数据集上的实验结果表明,该方法在缓解数据稀疏问题的同时,提高了推荐的效率和性能。

Graph convolutional neural networks have achieved great success in collaborative filtering recommendation systems.However,in real recommendation scenarios,collaborative filtering-based recommendation methods are often affected by sparse data,resulting in poor recommendation accuracy.Plus,existing graph collaborative filtering methods generally suffer from incomplete analysis and utilization of user-item interaction information,e.g.,fail to deal with noise in interaction features,and these lead to unsatisfactory recommendation results.To address the above problems,this paper proposes a graph collaborative filtering method that combines prototype comparison and feature filtering.It captures potential connections between nodes by using the proposed prototype contrastive learning task to enhance the representation of users and items,while removes noise from the interaction information.The results of the experiments on three real datasets show that the method improves the accuracy of recommendations while alleviating the data sparsity problem.