Experiment & Benchmark Paper: To What Extent Does Quality Matter? The Impact of Graph Data Quality on GNN Model Performance

Abstract

Real-world data often is noisy and error-prone which can negatively influence machine learning models. Graph Neural Networks (GNNs) introduce the additional challenge that during training, node information is iteratively passed through the graph along the edges. Consequently, errors or deviations in the graph data could highly impact the model’s predictive capability. Our work systematically investigates how quality deviations in graph datasets influence the GNN model performance. We focus on the node features and explore three dimensions: the rate of modified features, the amplitude of modification, and the feature precision. Based on our results, we give insights and recommendations for practitioners. For instance, when using highly clustered graphs, modifying around 40% of the features only results in a slight decrease of performance and the rate of modified features is more crucial than the amplitude of modification. We illustrate practical implications of our results by establishing connections to real world domains such as graph dataset acquisition and efficient GNN training.

License