SolePostHoc.jl – Post-Hoc Analysis for Symbolic Learning

In a nutshell

SolePostHoc.jl provides a comprehensive suite of post-hoc analysis and optimization tools for symbolic learning models. The package enables knowledge extraction from both symbolic and non-symbolic models through a unified interface, facilitating the comparison of different interpretation methods while maintaining consistency and ease of use.

Key Features

Knowledge Extraction Algorithms

• LUMEN;
• InTrees;
• TREPAN;
• REFNE;
• RuleCOSI+;
• BATrees.

Model Analysis Capabilities

• Rule extraction from decision trees, random forests, and black-box models
• Model transformation and enhancement through rule minimization
• Performance optimization while preserving interpretability
• Support for surrogate model generation and knowledge distillation

Unified Interface

All algorithms are accessible through a consistent API via the SolePostHoc.modalextractrules function, enabling seamless comparison between different extraction methods.

Usage

Direct Algorithm Access

# Call specific algorithms directly
extracted_rules = lumen(model)
extracted_rules = intrees(model, X_test, y_test)

Unified Interface (Recommended)

# Use the unified interface with rule extractors
extractor_lumen = LumenRuleExtractor()
extractor_intrees = IntreesRuleExtractor()

decision_set_lumen = modalextractrules(extractor_lumen, model)
decision_set_intrees = modalextractrules(extractor_intrees, model, X_test, y_test)

The unified interface converts outputs into DecisionSet objects—vectors of propositional logical rules in Disjunctive Normal Form (DNF), with one rule per class label.

Example: Rule Extraction from Random Forest

using SolePostHoc

# Assume we have a trained Random Forest on the Iris dataset
rf_model = train_random_forest(X_train, y_train)

# Extract interpretable rules using LUMEN
extractor = LumenRuleExtractor()
interpretable_rules = modalextractrules(extractor, rf_model)

# The result is a DecisionSet with logical rules explaining the model's decisions

Algorithm Categories

Surrogate Trees: Approximate complex models with interpretable decision trees.

Knowledge Distillation: Transfer knowledge from complex to transparent models.

Rule Extraction: Derive clear logical rules from any machine learning model.

Integration

SolePostHoc.jl is part of the Sole.jl ecosystem:

• SoleLogics.jl: Logical foundations
• SoleData.jl: Data handling
• SoleModels.jl: Model definitions
• SoleFeatures.jl: Feature engineering

References

For theoretical foundations, see: Modal Symbolic Learning: from theory to practice, G. Pagliarini (2024)

About

Developed by the ACLAI Lab @ University of Ferrara.