基于协同信号的知识图注意力网络推荐算法

Collaborative signal knowledge graph attention network for recommender algorithm

针对现有基于知识图谱的推荐算法大多采用随机采样的方法构建感受野导致部分重要信息丢失的问题,提出一种基于协同信号的知识图注意力网络推荐算法(CKGAN)。设计协作传播层,将用户的历史项目和待推荐项目在知识图谱中进行传播;为区分不同关系下的实体对中心节点的重要性不同,在聚合邻域信息时引入图注意力网络;为解决用户的不同兴趣,在预测层中,面对不同的候选集动态建模用户的向量表示。通过在Last.FM、Book-Cross、MovieLens-1M这3个真实的公共数据集上与主流的推荐算法进行对比实验,其结果表明,CKGAN在CTR预测和top-K推荐中都有显著提升。

The existing recommendation algorithm based on knowledge graph adopts the method of random sampling to construct the receptive field,which leads to the loss of some important information.Regarding the issues,a recommendation algorithm called collaborative signal knowledge graph attention network for recommender algorithm(CKGAN)was proposed.The collaborative propagation layer was designed to spread users’historical items and items to be recommended in KG.To distinguish the importance of entities under different relations to the central node,graph attention network was introduced in the propagation process.To address the different interests of users,in the prediction layer,the vector representation of users was dynamically modeled in the face of different candidate sets.Through comparative experiments with state-of-the-art algorithms on three real public datasets,Last.FM,Book-Cross and MovieLens-1M,the results show that CKGAN significantly improves both CTR prediction and top-K recommendation.