One-class graph autoencoder: A new end-to-end, low-dimensional, and interpretable approach for node classification

Abstract

One-class learning (OCL) for graph neural networks (GNNs) comprises a set of techniques applied when real-world problems are modeled through graphs and have a single class of interest. These methods may employ a two-step strategy: first representing the graph and then classifying its nodes. End-to-end methods learn the node representations while classifying the nodes in OCL process. We highlight three main gaps in this literature: (i) non-customized representations for OCL; (ii) the lack of constraints on hypersphere learning; and (iii) the lack of interpretability. This paper presents One-cLass Graph Autoencoder (OLGA), a new OCL for GNN approach. OLGA is an end-to-end method that learns low-dimensional representations for nodes while encapsulating interest nodes through a proposed and new hypersphere loss function. Furthermore, OLGA combines this new hypersphere loss with the graph autoencoder reconstruction loss to improve model learning. The reconstruction loss is a constraint to the sole use of the hypersphere loss that can bias the model to encapsulate all nodes. Finally, our low-dimensional representation makes the OLGA interpretable since we can visualize the representation learning at each epoch. OLGA achieved state-of-the-art results and outperformed six other methods with statistical significance while maintaining the learning process interpretability with its low-dimensional representations.