Graph neural network toward representation alignment and uniformity for session-based recommendation
Session-based recommendation(SBR)aims to predict the next click item of an anonymous user based on a short interaction sequence.In order to solve the problem that existing SBR methods based on graph neural networks(GNNs)ignore the differences between the same items at different locations in a session,after obtaining an item embedding by a GNN,we further consider the correlation of its neighbor items in a session to generate the item neighborhood correlation representation.Considering the importance of alignment and uniformity in contrastive learning,we propose an alignment and uniformity loss method for session recommendation to constrain the generated session representation and item representation.Experiments on three public benchmark datasets show that our newly proposed model TAU-GNN is better than that of the mainstream models for SBR.
session-based recommendationgraph neural networksalignmentuniformitycontrastive learningcross entropy lossanonymous sessionneighborhood information
国家科技期刊平台
NSTL
万方数据
维普
