Singular Value Decomposition for prevention of the Graph Convolutional Network overfitting

Abstract

Our project is an effort to prevent the graph neural network from being overfitted. Graph convolutional networks are usually shallow, with the number of layers not larger than 2. Deep graph networks perform much worse, even if some standard techniques like dropout and weight penalizing are being implemented. In our work, we use singular value decomposition(SVD) for extracting the most relevant components from embeddings constructed by the network to overcome overfitting.