Effects of Topological Factors and Class Imbalance on Node Classification Through Graph Convolutional Neural Networks

Graph Convolutional Neural Networks (GCNs) have proven to be highly effective in solving graph-related problems, as they not only consider the individual node features but also capture the topological characteristics of the graph. However, the lack of public datasets presents a challenge in the evaluation and comparison of these networks across various contexts. This article addresses the inherent limitations of GCNs, focusing specifically on the impact of topological aspects and class imbalance in node classification tasks. Using a variant of the Stochastic Block Model (SBM) algorithm that allows for node covariates, a statistically significant number of synthetic graphs is generated, varying feature characteristics as well as group edge probabilities. Thus, a comprehensive exploration of GCNs' capabilities in different scenarios is conducted. The initial findings underscore the fundamental importance of node feature variability for classification and highlight the challenges that arise when presented with strong class imbalance scenarios.