Extracting rules from neural networks as decision diagrams

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Neural networks (NNs) often function as "black boxes", limiting understanding of their decision-making processes.
• Rule extraction from NNs can provide insights into network logic and improve generalization capabilities.
• Existing methods for rule extraction may have limitations in complexity or generalization.

Purpose of the Study:

• To present a novel eclectic approach for rule extraction from multilayer perceptron networks.
• To develop a method that enhances the interpretability and generalization of neural network models.
• To introduce a technique suitable for neural networks with discrete input types.

Main Methods:

• A novel approach named LOcal Rule Extraction (LORE) is proposed.
• Multilayer perceptron networks are trained and then transformed into an equivalent form for rule extraction.
• Partial rules are merged using a decision diagram (DD) data structure to form generalized rules.
• A specific rule format is presented to handle known and undetermined input subsets.

Main Results:

• The LORE method produces rules that mimic network behavior on training data and generalize to unseen data.
• Extracted rules demonstrate comparable complexity and generalization abilities to existing methods.
• The rule extraction process, including decision diagram transformation, is computationally efficient, with running times proportional to network training time.

Conclusions:

• LORE offers an effective method for extracting interpretable rules from neural networks with discrete inputs.
• The approach balances mimicking training data behavior with generalizing to new inputs.
• The use of decision diagrams facilitates efficient rule merging and extraction, contributing to improved AI model transparency.