PLEX

Official Paper: https://doi.org/10.1007/978-3-032-37664-0_28

A self-explainable Graph Neural Network architecture that uses Polyadic Graded Modal Logic (PGML) to support edge-aware reasoning via ternary predicates R(x, y, e). Supports graph classification, node classification, and link prediction tasks.

Repository Structure

Models

File	Description
model.py	Teacher (Black-box) and logic (PLEX) models for graph classification
model_node.py	Models for node classification
model_link.py	Models for link prediction

Training

File	Description
train_baseline.py	Train teacher model (graph classification)
train_baseline_node.py	Train teacher model (node classification)
train_baseline_link.py	Train teacher model (link prediction)
train_logic.py	Train PLEX (graph classification)
train_logic_node.py	Train PLEX (node classification)
train_logic_link.py	Train PLEX (link prediction)

Hyperparameter Optimization

File	Description
optimize_baseline.py	Hyperparameter search for teacher models on graph classification
optimize_baseline_node.py	Hyperparameter search for teacher models on node classification
optimize_baseline_link.py	Hyperparameter search for teacher models on link prediction
optimize_logic.py	Hyperparameter search for PLEX on graph classififcation
optimize_logic_node.py	Hyperparameter search for PLEX on node classification
optimize_logic_link.py	Hyperparameter search for PLEX on link prediction

Explainability & Analysis

File	Description
LayerWiseRules(GC).ipynb	Extract logic rules from trained model (graph classification)
LayerWiseRules(NC).ipynb	Extract logic rules (node classification)
LayerWiseRules(LP).ipynb	Extract logic rules (link prediction)
RuleActivations.py	Visualize rule activations
HoyerRuleStatistics.py	Hoyer sparsity statistics and plots

For the dataset Aromatic-Carbon there are different files for training, optimization and explainability.