BAED: A new paradigm for few-shot graph learning with explanation in the loop

Area of Science:

• Graph Representation Learning
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot graph learning (FSGL) faces challenges with insufficient labeled data due to expert annotation requirements.
• Existing FSGL methods may sacrifice robustness and interpretability, leading to overfitting and performance degradation.

Purpose of the Study:

• Introduce the first explanation-in-the-loop framework for FSGL, named BAED.
• Improve model adaptability, robustness, and interpretability in few-shot graph learning scenarios.

Main Methods:

• Employ belief propagation for label augmentation on graph data.
• Utilize an auxiliary graph neural network and gradient backpropagation to extract explanatory subgraphs.
• Base final predictions on informative subgraphs to mitigate noise from neighboring nodes.

Main Results:

• BAED demonstrates superior prediction accuracy across seven benchmark datasets.
• The framework achieves enhanced training efficiency compared to existing methods.
• BAED provides high-quality explanations, improving the interpretability of FSGL models.

Conclusions:

• BAED represents a significant advancement in FSGL by integrating explanation mechanisms.
• The explanation-in-the-loop paradigm shows strong potential for addressing FSGL challenges.
• This work paves the way for more robust, interpretable, and efficient few-shot graph learning systems.