Session-based recommendation with graph neural networks for repeat consumption

Session-based recommendation aims to predict user actions only based on anonymous session sequence. The repeat consumption is a common phenomenon in the recommendation scenarios which the items will be clicked repeatedly many times. Existing recommended methods for session-based repeat consumption model a session as sequence and use Markov Chain (MC) or Recurrent Neural Network (RNNs) to generate the representations of items. Although achieved promising results, these models still have the deficiencies in obtaining the correct user vectors and complex transitions of items. On the other hand, the number of unclicked items is very large relative to the items that have been clicked in a session, that it is more difficult to mine information for unclicked items. In this paper, we proposed an improved model named GNN-RepeatNet based on RepeatNet to explicitly model the repeat consumptions in session-based recommendation, by utilizing the graph neural network and multi-layer self-attention. In GNN-RepeatNet, we get accurate item embedding and complex transitions of items via graph neural network (GNN). Then through a repeat-explore mechanism, we compute the probabilities of predicted items in repeat mode and deep-explore mode separately. In addition, the multi-layer self-attention networks is introduced to deeply explore the unclicked items in deep-explore model. Extensive experiments conducted on two real datasets show that GNN-RepeatNet can improve the performance compared to the state-of-the-art methods.