Identify Spammers in Rating Systems Using Multi-layer Graph Convolutional Network

Abstract

With the rise of deep learning technology, researches on the application of graph convolutional networks to fraud detection emerge endlessly. However, the graph convolutional network used often has only two layers, which makes us unable to obtain higher-order node information. So we must use a multi-layer graph neural network to get more node information for training in order to get more accurate detection results. Using a multi-layer graph neural network also causes gradient disappearance; that is, the model parameters cannot be updated, and the model is invalid. This work explores the feasibility of using multi-layer GCN to detect fraudsters from the internal structure of GCN, that is, the number of hidden layer neurons and the activation function. Finally, it is tested on real data sets, and the detection accuracy of fraudsters using multi-layer GCN is increased by about 14.6% at most.