Uncertainty-Aware Graph Neural Networks: A Multihop Evidence Fusion Approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph neural networks (GNNs) are powerful for graph representation learning but struggle with prediction uncertainty.
• Uncertainty in class probabilities increases with model depth, leading to unreliable predictions in real-world applications.

Purpose of the Study:

• To propose a novel evidence-fusing graph neural network (EFGNN) for trustworthy predictions and enhanced node classification.
• To explicitly quantify the risk of incorrect predictions by integrating evidence theory with GNNs.

Main Methods:

• Developed an evidence-fusing graph neural network (EFGNN) integrating evidence theory with multihop GNNs.
• Quantified node prediction uncertainty by considering multiple receptive fields.
• Introduced a parameter-free cumulative belief fusion (CBF) mechanism to enhance prediction trustworthiness.
• Designed a joint learning objective including evidence cross-entropy, dissonance coefficient, and false confident penalty.

Main Results:

• Demonstrated the effectiveness of EFGNN in improving node classification accuracy.
• Showcased EFGNN's ability to provide trustworthy predictions by quantifying uncertainty.
• Validated the model's robustness against potential attacks through theoretical analyses and experiments.

Conclusions:

• The proposed EFGNN effectively addresses the uncertainty issue in GNNs, leading to more reliable predictions.
• EFGNN enhances node classification accuracy and provides explicit risk assessment for predictions.
• The model offers a robust solution for trustworthy AI in graph-based learning scenarios.