PL4XGL: A Programming Language Approach to Explainable Graph Learning
summary
Proposes PL4XGL, a GDL-based classifier that learns models by program synthesis (top-down and bottom-up) and returns the synthesised GDL program as a provably correct explanation with no extra posthoc cost. Reports competitive accuracy with popular GNNs on multiple node and graph datasets, outperforms SubgraphX on explanation quality (better tradeoff of fidelity and sparsity), trains on CPU while baselines use GPU, but times out on the largest Hive set of 40,000 graphs (1M nodes). GDL programs drive predictions.
AI-written summary · about
chapters (10)
- 0:00PL4XGL and explainable graph learning
- 1:50Limitations of GNN explanations
- 3:35Designing explainable graph ML
- 4:57Model example walkthrough
- 8:57Program synthesis for training
- 9:39Evaluation and limitations
- 11:48Key contributions
- 12:53Q&A: When does this method work best?
- 13:34Q&A: Relation to graph query languages
- 14:37Closing remarks
related talks
26mDeep Graph Library: Deep Graph learning at scale
15mGauss: Program Synthesis by Reasoning Over Graphs
31mGraph-Based Machine Learning with Python
15mGALA: A High Performance Graph Neural Network Acceleration LAnguage and Compiler
28mExplaining Machine Learning Models with LLMs
29mGraph Neural Networks + LLM on
35mScalable structure learning and inference for domain-specific probabilistic prog
01h 01mML for discrete optimization: Theoretical foundations