PEVGraphRec: a PEV method-based graph neural networks for social recommendations

Abstract

Most existing recommendation systems assume that items are independent and identically distributed. These methods only leverage user-item interactions as model input, while neglecting the relationship between items, leading to limited performance. To address this issue, this paper proposes the PEVGraphRec, which provides recommendations by utilizing Proximity-Effect-Value (PEV) similarity measurement to generate an item-item implicit network and employing graph neural network with attention mechanisms to intelligently aggregate the user-user connections, the user-item interactions, and the item-item relations. Specially, compared with traditional similarity measurement, the PEV method can reduce the negative effects of cold start. Experiments on three real-world datasets verify the interpretability of our model by examining the contribution of each component. The results show that the PEVGraphRec is robust and superior to other baseline methods.