卷积记忆图协同过滤

针对推荐系统中用户和项目的向量表示问题，提出了一种端到端的具有记忆单元的图神经网络.在图神经网络中引入门控循环单元解决高阶连通节点间信息损失问题，可以使得用户和项目节点从高阶邻居获得更加完整的特征信息，然后利用卷积神经网络对网络输出层间的特征向量进行融合以获得不同阶段下用户的偏好.实验结果表明，与最优对比算法相比，采用所提卷积记忆图协同过滤推荐算法在4个数据集上的评分预测性能分别提升了1.98%，4.17%，9.27%和2.70%.

Convolutional Memory Graph Collaborative Filtering

An end-to-end graph neural networks with memory unit is proposed for user vector representations and items in recommender systems. Gated recurrent unit is introduced to reduce the information loss between high-order connected nodes. This enables users and items nodes to obtain more complete feature information from high-order neighbor nodes. The convolutional neural networks are used to fuse feature vectors between different output layers to obtain users' preferences at different stages. Experiments on 4 datasets show that compared with the optimal comparison algorithms, the performance of proposed algorithm achieves gain of 1.98%, 4.17%, 9.27% and 2.7%, respectively.