Evidential Self-Supervised Graph Representation Learning via Prototype-based Consistency

This paper investigates self-supervised graph representation learning, addressing the challenges posed by noise and ambiguity inherent in graph data, which often result in low data quality. To enhance the trustworthiness of learned graph representations, this paper proposes a novel framework. It first employs a graph encoder to capture both local and global information from the original graph and its augmented versions produced through graph diffusion techniques. To better mitigate the effects of data ambiguity, our framework constructs learnable prototypes to cover the entire semantic space and utilizes Subjective Logic to provide evidence and quantify the uncertainty in graph data, thereby learning trustworthy and high-quality graph representations.