图神经网络具备融合节点信息与拓扑结构的能力,近年来在推荐算法中得到了广泛的应用.然而,现有的基于图神经网络的推荐模型用户行为建模粒度较粗,用户特征学习算法对历史信息使用不足,两者阻碍了用户偏好特征的提取.针对以上问题,本文提出一种基于增强多通道图注意力的推荐模型(en-hanced multi-channel graph attention based collaborative filtering recommendation model,EMGACF).在邻域聚合部分,采用多通道图注意力对细粒度用户评分等级建模,有效提升了模型对用户偏好的学习能力;在节点更新部分,提出基于增强自信息的节点更新算法,使用邻居节点聚合表示的同时保留了节点自身历史信息和内在偏好,提升了迭代过程中用户偏好的学习效果.实验部分在4种规模的常用推荐系统基准数据集上训练模型,实验结果表明,预测误差相比于主流模型降低了1.43%~7.81%.
Enhanced multi-channel graph attention based on recommendation model
Graph neural networks can fuse node information and topology,and are widely used in recommendation algorithms in recent years.However,the existing recommendation models based on graph neural networks have coarse granularity when modeling user behavior,and the user feature learning algorithm lacks consideration of historical information,both of which hinder the extraction of user preference features.To address the above problems,this paper proposes an enhanced multi-channel graph attention based collaborative filtering recommendation model(EMGACF).In the neighborhood aggregation phase,multi-channel graph attention is used to model fine-grained user rating levels,which effectively improves the learning ability of the model for user preferences;in the node update phase,an enhanced self-information-based node update algorithm is proposed,which uses the aggregated representation of neighboring nodes while preserving the nodes'own historical information and intrinsic preferences.This improves the learning effect of user preferences in the iterative process.By training the model on three benchmark datasets of different sizes of recommender systems,the experimental results show that the prediction error is reduced by 1.43%to 7.81%compared with the mainstream model.