AGLGCN: Graph Convolution Recommendation Algorithm based on Multi-Head Attention and Gated Recurrent Unit

Abstract

To solve the problems that LightGCN allocates the same weight to each user (item) and does not integrate the time feature, this paper puts forward a new graph convolution recommendation algorithm model(AGLGCN) based on the LightGCN with multi-head attention and gated recurrent unit. Firstly, the interactive feature of users and items is obtained in the data set and converted into feature vectors. Secondly, the feature vectors are introduced into the graph convolution neural network and aggregated. Then it was introduced into multi-head attention and GRU neural network to assign different weights to different features and learn the time feature information through the gated recurrent unit. Finally, BPR parameters were used to optimize, and the final recommendation results were obtained. In this paper, MF, NeuMF, GCMC, NGCF, and LightGCN models were recurred and compared with AGLGCN in two public data sets, MovieLens and Yelp. The four evaluation indexes of Recall@10, Recall@20, NDCG@10, and NDCG@20 were all improved by our proposed algorithm.