Beyond Sparse Benchmarks: Evaluating GNNs with Realistic Missing Features

Handling missing node features is a critical challenge for deploying Graph Neural Networks (GNNs) in real-world applications such as healthcare and sensor networks. This has led to a number of recent works exploring techniques for learning GNNs from incomplete data. However, existing evaluations are often based on benchmark datasets with high-dimensional but very sparse node features, where predictive performance degrades only slowly as the proportion of missing values increases. In this paper we move towards more challenging and realistic scenarios by considering datasets in which the predictive signal is more sensitive to feature incompleteness. We provide a theoretical background for clearly identifying relevant assumptions on the missingness mechanism, and for analyzing their implications for different solution approaches. Based on this analysis, we introduce the GNNmim approach for node classification in graphs with incomplete feature data. Experiments show that GNNmim consistently outperforms more complex models across a range of datasets and levels of missingness.