Social Recommendation based on Graph Neural Networks

Social recommendation has become a hot research topic in recommendation applications. However, current studies have not well simultaneously portrayed the static and dynamic features of users and items. Especially, they do not combine time information, social relations, and users' previous rating information which are of vital importance in learning the feature representations of users and items during modeling. Besides, graph neural networks that integrate social relations into the feature representation of users and items have also achieved superior performance. Thus, we jointly consider all the above information and thereby propose a dynamic social recommendation model named DGARec-R by adopting attention-based graph convolutional networks and recurrent neural networks. Experimental results on two real-world datasets show that our model not only outperforms baseline methods and its variations, but also is good at dealing with cold-start users.