Extraction of the association rules from artificial neural networks based on the multiobjective optimization

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Artificial Neural Networks (ANNs) are powerful machine learning tools for prediction and classification.
• A significant barrier to ANN adoption is their lack of transparency, often termed the 'black box' problem.
• This opacity hinders trust and validation in critical applications.

Purpose of the Study:

• To develop a novel rule extraction algorithm for ANNs.
• The algorithm aims to generate comprehensible rules that accurately mimic ANN decisions.
• The goal is to achieve a balance between rule fidelity, accuracy, and comprehensibility.

Main Methods:

• The proposed algorithm involves three phases: ANN learning, rule extraction, and rule simplification.
• Rule extraction utilizes association rule mining techniques.
• Rule simplification employs principles of Boolean algebra.

Main Results:

• The algorithm was evaluated on four datasets and compared against existing rule extraction methods.
• The proposed method generated a concise set of rules.
• These rules demonstrated superior accuracy and fidelity compared to other approaches.

Conclusions:

• The developed algorithm effectively addresses the transparency issue in ANNs.
• It provides a method for generating accurate, high-fidelity, and comprehensible rules.
• This enhances the interpretability and trustworthiness of machine learning models.