Transaction-aware heterogeneous graph embedding for recommendation

Abstract

The auxiliary information describing users and items are widely used in the model of recommendation increasingly.Heterogeneous graph,as a effective means to incorporate these information, has been widely used in the modelling of the auxiliary information of the users and items.Existing models usually fail to capture relevance of user and its high-order neighbors,likewise the item.Besides,existing models represent the user without considering the effect of predicted item.To address the above issues,we encode high-order semantic relationships into user and item representations by information propagation along the graph.Besides,we design co-attention neural network to generate the transaction-aware embedding of both user and item to better consider the impact of different items to users.In all,we propose a transaction-aware heterogeneous graph embedding for recommendation(TA-HGRec).Experimental with thress real datasets showed that it achieved significant improvement over existing state-of-the-art recommendation methods.