【摘要】 目前，基于知识图谱的推荐方法利用图神经网络捕获用户偏好和知识实体之间的相关性，实现最优的推荐效果.但这种相关性建模方法依赖于节点(用户、物品或实体)之间的显式关系，具有一定的局限性.针对上述问题，文中提出隐式知识图协同过滤模型(Implicit Knowledge Graph Collaborative Filtering Model, IKGCF).首先，构建隐式协同知识图，一方面消除显式关系对推荐中隐式交互关系的干扰，另一方面解除显式关系对图谱中语义相关性的限制.然后，采用增强的图神经网络模块，执行邻居聚集和消息传播，更好地捕获隐式协同知识图上的高阶相关性.最后，采用层选择机制，得到最终的节点嵌入向量，并对模型进行预测和优化.在3个公开数据集上的实验表明，文中模型推荐效果较优.IKGCF的完整代码开源在https://github.com/hfutmars/IKGCF.更多还原

【Abstract】 In the existing recommendation methods based on knowledge graphs, graph neural networks are utilized to capture the correlation between user preferences and knowledge entities to achieve optimal recommendation results. However, certain limitations occur in this kind of relevance modeling methods due to its dependence on the explicit relationship between nodes(users, items or entities). To address these problems, an implicit knowledge graph collaborative filtering model(IKGCF) is proposed. Firstly, the implicit collaborative knowledge graph is constructed to eliminate the interference of explicit relationship on implicit interaction in recommendations and remove the limitation of explicit relationship on semantic relevance in the graph. Then, an enhanced graph neural network module is adopted to perform neighbor aggregation and message propagation to better capture the higher-order relevance on the implicit collaborative knowledge graph. Finally, a layer selection mechanism is employed to obtain the final node embedding vectors and predict and optimize the model. Experiments on three public datasets show that IKGCF achieves better performance. The full code of IKGCF is open-sourced at https://github.com/hfutmars/IKGCF.更多还原