Epistemic Graph Neural Networks

Abstract

This paper introduces Epistemic Graph Neural Networks (EGNNs), a novel approach to graph neural networks that explicitly models and incorporates uncertainty into the learning process. Traditional GNNs primarily focus on learning pattern recognition from graph-structured data, often neglecting the inherent uncertainty associated with these patterns and the relationships between nodes. EGNNs address this limitation by representing the network's confidence in inferred relationships as dynamic edge weights within a graph. Nodes in the graph represent individual data points, and edges represent the strength of the inferred relationships between them. Crucially, the weight of each edge is modulated by a learned epistemic uncertainty estimate, reflecting the network's confidence in that particular connection. This allows EGNNs to not only learn patterns but also to represent and reason with uncertainty, leading to more robust and reliable predictions and decision-making. The core claim of this work is that neural networks can learn not just patterns in data, but also the *uncertainty* inherent in those patterns, represented as a dynamically evolving graph. The mechanism is a GNN where nodes represent data points and edges represent the strength of inferred relationships, with edge weights modulated by learned epistemic uncertainty estimates – representing the network's confidence in the relationship.