Improved knowledge Graph Convolutional neural Network recommendation model

Abstract In the KG-aware recommendation system, only a few methods capture valuable information from both user–item interaction and item knowledge, and they introduce many trainable parameters that increase training difficulty. To solve the above problems, a simplified knowledge graph-aware & graph convolutional network recommendation model (SKGCN) was proposed, Firstly, feature transformation and nonlinear activation in Graph Convolution Layer are removed, which has been proven to greatly reduce parameters. Then, two specific aggregation components are introduced to the feature aggregation process of user-item interaction graph and item knowledge graph. Finally, the obtained embeddings of users and items were integrated to generate the prediction score for recommendation. The experimental results on the three datasets of MovieLens-20M, Book-Crossing and LastFM show that the Recall rate is improved by an average of 5.00%, 24.32% and 1.05%, and the normalized damage cumulative gain increased by an average of 2.17%, 23.71% and 1.71%, respectively, compared with optimal benchmark model.