多通道图注意力解耦社交推荐方法

提出具有解耦能力的多通道图注意力社交推荐模型,该模型主要包括深度聚类模块、多通道图注意力聚合模块和评分预测模块。其中,深度聚类模块用于对用户和项目进行分组,并利用聚类结果将用户社交图和用户项目图拆分成多个用户社交子图及用户项目子图,以学习用户兴趣分组及用户对不同类项目的兴趣;多通道图注意力聚合模块学习不同子图对预测结果的注意力;评分预测模块将学习到的用户表示向量和项目表示向量输入多层感知机进行评分预测。在多个真实数据集上的实验结果表明:提出的方法优于其他社交推荐算法。与最新的用于社交推荐的图神经网络方法相比,在Ciao和Epinions数据集上,均方根误差分别降低了2.26%和2.07%,平均绝对误差分别降低了2.58%和3.06%。

Multi-channel graph attention network with disentangling capability for social recommendation

In recent years, graph neural networks have proved to be very powerful in learning graph structure data since social networks and user rating information in the social recommendation dataset can be represented as user-user social graph and user-item interaction graph. Social recommendation methods based on graph neural networks have gradually become a research hotspot in recommendation field. Nowadays, social recommendation models based on graph neural network are facing with the following challenges. 1) How to effectively extract and integrate user-user social information and user-item interactive information. 2) How to distinguish the users" purchase motivations. Therefore, in this study, we propose a Multi-channel Graph ATtention network with Disentangling capability for Social Recommendation named MGAT-D for social recommendation. This model mainly includes three modules: deep clustering module, the aggregation module based on multi-channel graph attention network, and the rating prediction module. Among them, the deep clustering module is used to group users and items. The clustering results can then be used to split user-user social graph and user-item interaction graph into multiple subgraphs to capture user interest groups and users" interests in different types of items. The aggregation module learns the attention of different sub-graphs related to prediction results. The rating prediction module merges user embeddings and item embeddings into the multi-layer perceptron (MLP) to predict the ratings. Extensive experiments on multiple real-world datasets demonstrate that the proposed model MGAN-D is better than other social recommendation algorithms. Specifically, compared with the latest state-of-the-art model GraphRec (Graph Neural Networks for Social Recommendation), the decrease of RMSE (Root Mean Square Error) on the Ciao and Epinions datasets is 2.58% and 3.06%, and that of MAE (Mean Absolute Error) is 2.26% and 2.07%, respectively.