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 collabo-rative 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-1 M,the results show that CKGAN significantly improves both CTR pre-diction and top-K recommendation.
关键词
知识图谱/随机采样/感受野/图注意力网络/异构网络/隐式反馈/推荐系统
Key words
knowledge graph/random sampling/receptive field/graph attention network/heterogeneous network/implicit feed-back/recommendation system
维普
万方数据