G-MLP: Graph Multi-Layer Perceptron for Node Classification Using Contrastive Learning

Graph Convolutional Network (GCN) and its variants emerged as powerful graph deep learning methods with promising performance on graph analysis tasks. Different variants improve performance by introducing efficient information propagation and aggregation modules of GCN. To simplify the message passing modules, we propose the Graph Multi-Layer Perceptron (G-MLP), an innovative Multi-Layer Perceptron (MLP) method that uses contrastive learning to implicitly extract the original graph features and learn discriminative node representations. Firstly, we concatenate the topology and attributes to generate the feature matrix, and take it as the input. Secondly, linear layers in the MLP-based structure are combined with activation function and dropout to enhance the nonlinear capacity and prevent overfitting. Finally, we use a novel contrastive loss to optimize node representations, preserving node similarity in the feature space and bridging the gap between GCN and MLP. Experiments with the proposed method are conducted on node classification using four benchmark datasets. The results of G-MLP achieve state-of-the-art performance compared to other baselines, demonstrating that the contrastive loss can improve the representation capability of the MLP-based structure.