ANN-DT: an algorithm for extraction of decision trees from artificial neural networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Science

Background:

• Artificial neural networks (ANNs) offer high accuracy for complex systems but lack interpretability.
• Limited comprehensibility hinders ANN adoption in critical applications requiring reliability.
• Existing methods for extracting insights from ANNs are often restrictive.

Purpose of the Study:

• To introduce a novel artificial neural network decision tree (ANN-DT) algorithm.
• To enable the extraction of binary decision trees from trained neural networks.
• To overcome limitations of existing techniques, particularly for feedforward networks with continuous outputs.

Main Methods:

• The ANN-DT algorithm generates outputs for interpolated samples using the trained neural network.
• It extracts decision rules without assumptions on network architecture or data features.
• A new attribute selection criterion based on significance analysis is evaluated.

Main Results:

• ANN-DT successfully extracts rules from feedforward neural networks with continuous outputs.
• The novel attribute selection criterion shows benefits over standard variance-based methods.
• In three case studies, ANN-DT demonstrated favorable performance compared to CART.

Conclusions:

• The ANN-DT algorithm provides a viable method for enhancing the interpretability of artificial neural networks.
• This approach expands the practical applicability of ANNs in domains demanding transparency.
• ANN-DT offers a competitive alternative to traditional decision tree algorithms.