Sensitivity to Noise in Features in Graph Neural Network Learning

Graph Neural Networks learn with two types of information: the structure of the graph and the features of the nodes and edges. These features can be noisy or incomplete. For instance, some social network users are more active than others, and thus produce more information. In this paper, we explore the performance of Graph Neural Networks when the graph contains noisy features. We train three architectures that represent the basis of many of today’s state-of-the-art methods. We use an academic citation network in which we manually add noise to the features. We investigate three types of noise: random noise, noise that targets the influential nodes, and noise that targets the most isolated nodes. We train each model in both transductive and inductive settings, i.e., when the model has access to test data at training time, or when the test data is hidden. We find that the models are most vulnerable to noise targeting the isolated nodes. Additionally, we empirically validate that some models present implicit denoising abilities. We also find that the lack of regularizing mechanisms, such as weighting schemes, negatively impact the performance of the models.