PL4XGL: A Programming Language Approach to Explainable Graph Learning

Explore a novel programming language approach to explainable graph learning in this 15-minute conference talk from PLDI 2024. Discover PL4XGL, a new graph-learning method that addresses limitations in explainability of graph neural networks (GNNs). Learn about the innovative graph description language (GDL) designed to explain classification results and the development of a GDL-based interpretable classification model. Understand how this approach formulates learning from data as a program-synthesis problem, with top-down and bottom-up algorithms for synthesizing GDL programs from training data. Examine the evaluation results demonstrating PL4XGL's ability to produce high-quality explanations that outperform state-of-the-art GNN explanation techniques while achieving competitive classification accuracy. Gain insights into this groundbreaking research that combines graph learning, domain-specific languages, and program synthesis to enhance explainability in decision-critical applications.