基于深度图神经网络的协同推荐算法

针对现有基于图神经网络(GNN)的推荐算法面临的过平滑的问题,提出一种基于深度GNN的协同过滤推荐算法Deep NGCF(Deep Neural Graph Collaborative Filtering)。该算法在GNN中引入初始残差连接和恒等映射,避免了GNN进行多次图卷积运算后陷入过平滑。首先,通过用户和项目的交互历史得到它们的初始嵌入;其次,在聚合传播层利用初始残差连接和恒等映射得到用户和项目的不同阶协同信号;最后,对所有协同信号进行线性表示以得到预测评分。此外,在初始残差连接和恒等映射中设置比重进行调节,从而进一步提高模型的灵活性和推荐性能。为验证Deep NGCF算法的可行性和有效性,在Gowalla、Yelp-2018与Amazon-Book数据集上进行实验。实验结果表明,相较于图卷积矩阵补全(GCMC)、神经图协同过滤(NGCF)等现有的GNN推荐算法,Deep NGCF算法取得了最高的召回率和归一化折损累计增益(NDCG),验证了所提算法的有效性。

Thermopower of YBa_(2)Cu_(3)O_(7) Superconductor Prepared by Melt-Textured Growth Method

The results of thermopower measurements in a high J_(c) melt-textured growth(MTG)YBa_(2)Cu_(3)O_(7) bulk sample above 80K are reported,its behavior is strongly anisotropic.The small value in c-direction indicates the existence of an imperfection in the alignment of ab-planes and low resistance channels along the c-axis,which may be important for the high Jc value of the sample.Described by a two order tensor,the orientational dependence of thermopower can also be explained.