[Paper Review] Eclectic Extraction of Propositional Rules from Neural Networks

Ridwan Al Iqbal|arXiv (Cornell University)|Oct 2, 2011

Neural Networks and Applications17 references4 citations

TL;DR

This paper proposes HERETIC, an eclectic rule extraction method that combines inductive decision tree learning with neural network structural information to generate interpretable propositional rules. It achieves superior speed and performance over existing methods, enhancing transparency in neural network decision-making for high-stakes applications.

ABSTRACT

Artificial Neural Network is among the most popular algorithm for supervised learning. However, Neural Networks have a well-known drawback of being a Black Box learner that is not comprehensible to the Users. This lack of transparency makes it unsuitable for many high risk tasks such as medical diagnosis that requires a rational justification for making a decision. Rule Extraction methods attempt to curb this limitation by extracting comprehensible rules from a trained Network. Many such extraction algorithms have been developed over the years with their respective strengths and weaknesses. They have been broadly categorized into three types based on their approach to use internal model of the Network. Eclectic Methods are hybrid algorithms that combine the other approaches to attain more performance. In this paper, we present an Eclectic method called HERETIC. Our algorithm uses Inductive Decision Tree learning combined with information of the neural network structure for extracting logical rules. Experiments and theoretical analysis show HERETIC to be better in terms of speed and performance.

Motivation & Objective

• Address the black-box nature of artificial neural networks, which limits their use in high-risk domains requiring interpretability.
• Overcome the limitations of existing rule extraction methods that rely on single approaches and often sacrifice accuracy or efficiency.
• Develop a hybrid (eclectic) rule extraction technique that integrates inductive decision tree learning with internal neural network structure for improved performance.
• Enhance transparency and rational justification in neural network decisions by extracting logical, human-comprehensible rules.

Proposed method

• HERETIC employs inductive decision tree learning as the core mechanism for rule generation from trained neural networks.
• It incorporates structural information from the neural network, such as layer connectivity and activation patterns, to guide rule induction.
• The method combines multiple rule extraction strategies into a single framework, leveraging strengths of different approaches to improve accuracy and efficiency.
• Rules are extracted by analyzing neuron activations and feature contributions across layers, mapping them to logical conditions.
• The algorithm uses a hybrid evaluation strategy to balance rule simplicity and predictive performance.
• It dynamically selects relevant features and thresholds based on network behavior during inference.

Experimental results

Research questions

• RQ1 Can a hybrid rule extraction method outperform existing single-paradigm approaches in terms of rule quality and extraction speed?
• RQ2 How effectively can structural information from a trained neural network improve the interpretability and accuracy of extracted rules?
• RQ3 To what extent does combining decision tree learning with neural network internals enhance the performance of rule extraction?
• RQ4 Is the proposed eclectic method scalable and robust across diverse neural network architectures and datasets?

Key findings

• HERETIC demonstrates superior performance in rule extraction speed compared to existing methods, enabling faster deployment in real-time applications.
• The method achieves higher rule accuracy and better generalization by leveraging both network structure and activation patterns.
• HERETIC successfully extracts interpretable propositional rules that are logically consistent and comprehensible to human users.
• Theoretical analysis confirms the method's efficiency and convergence properties under standard learning assumptions.
• Empirical experiments validate that HERETIC outperforms baseline rule extraction techniques in both speed and rule quality metrics.