异质信息网络高阶层次化嵌入学习与推荐预测

异质信息网络是一种异质数据表示形式,如何融合异质数据复杂语义信息,是推荐系统面临的挑战之一.利用弱关系具有的丰富语义和信息传递能力,构建一种面向推荐系统的异质信息网络高阶嵌入学习框架,主要包括:初始化信息嵌入、高阶信息嵌入聚合与推荐预测3个模块.初始化信息嵌入模块首先采用基于弱关系的异质信息网络最佳信任路径筛选算法,有效地避免在全关系异质信息网络中,采样固定数量邻居造成的信息损失,其次利用新定义的基于多头图注意力的多任务共享特征重要性度量因子,筛选出节点的语义信息,并结合交互结构,有效地表征网络节点;高阶信息嵌入聚合模块通过融入弱关系及网络嵌入对知识良好的表征能力,实现高阶信息表达,并利用异质信息网络的层级传播机制,将被采样节点的特征聚合到待预测节点;推荐预测模块利用高阶信息的影响力推荐方法,实现了推荐任务.该框架具有嵌入节点类型丰富、融合共享属性和隐式交互信息等特点.最后,实验验证UI-HEHo学习框架可有效地改善评级预测的准确性,以及推荐生成的针对性、新颖性和多样性,尤其是在数据稀疏的应用场景中,具有良好的推荐效果.

Higher-order Hierarchical Embedding Learning and Recommendation Prediction in HIN

Heterogeneous information network is a representation of heterogeneous data. How to integrate complex semantic information of heterogeneous data is one of the challenges faced by recommendation systems. A higher-order embedded learning framework for heterogeneous information networks based on weak ties featured by semantic information and information transmission abilities is constructed. The framework includes three modules of initial information embedding, high-order information embedding aggregation, and recommendation prediction. The initial information embedding module first adopts the best trust path selection algorithm to avoid information loss caused by sampling a fixed number of neighbors in a full-relational heterogeneous information network. Then the newly defined importance measure factors of multi-task shared characteristics based on multi-head attention are adopted to filter out the semantic information of each node. Additionally, combined with the interactive structure, the network nodes are effectively characterized. The high-order information embedding aggregation module realizes high-order information expression by integrating weak ties and good knowledge representation ability of network embedding. The hierarchical propagation mechanism of heterogeneous information networks is utilized to aggregate the characteristics of sampled nodes into the nodes to be predicted. The recommendation prediction module employs the influence recommendation method of high-order information to complete the recommendation. The framework is characterized by rich embedded nodes, fusion of shared attributes, and implicit interactive information. Finally, the experiments have verified that UI-HEHo can effectively improve the accuracy of rating prediction, as well as the pertinence, novelty and diversity of recommendation generation. Especially in application scenarios with sparse data, UI-HEHo yields good recommendation effects.