融合知识图卷积网络的双端邻居推荐算法

A Tow-endian Neighbor Recommendation Algorithm for Convolutional Networks Fused with Knowledge Graph

针对现有的基于知识图谱的推荐对于用户信息的考虑少于对物品信息的考虑,提出一种融合知识图卷积网络的双端邻居推荐算法,在用户端及物品端同时进行特征提取。对于用户特征的提取,是通过用户偏好在知识图谱中的扩散过程实现。对于物品特征的提取,是将邻居信息聚合到物品节点生成嵌入向量,因各个邻居的权重与用户点击物品的邻居节点紧密联系,因此基于KGCN模型来实现。最后让用户兴趣传播与物品特征聚合交替进行。在两个数据集上进行对比实验,在MovieLens-1M数据集上,与基线方法相比,AUC和F1分别提升了1.5%和2.0%,在Book-Crossing数据集上,AUC和F1分别提升了5.3%和1.9%,算法有效性得到显著提升。

Aiming at the fact that the existing recommendation based on knowledge graph pays less attention to user information than item information,a two-end neighbor recommendation algorithm based on knowledge graph convolution network is proposed to extract features at both user and item sides.The extraction of user personalized features is carried out through the diffusion process of user preferences in the knowledge graph.For item feature extraction,neighbor information is aggregated to item node to generate embedding vector.Since the weight of each neighbor is closely related to the neighbor node of the item clicked by the user,the vector of item feature is extracted based on KGCN model.Finally,user interest dissemination and item feature aggregation are carried out in turn.Comparison experiments were conducted on two data sets.Compared with baseline method,AUC and F1 improved by 1.5%and 2.0%respectively on Movielens-1M data set,while AUC and F1 improved by 5.3%and 1.9%respectively on Book-Crossing data set.The effectiveness of the proposed algorithm has been significantly improved.