Minerva: sequential covering for rule extraction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Advanced machine learning models like artificial neural networks and support vector machines offer superior performance.
• A key limitation of these models is their lack of interpretability, hindering understanding of their decision-making processes.
• Rule extraction (RE) techniques aim to address this opacity by generating understandable rules.

Purpose of the Study:

• To introduce Minerva, a novel algorithm for rule extraction.
• To overcome the limitations of existing RE techniques, which are often restricted to specific model types or data formats.
• To enable rule extraction from any black-box model, regardless of its complexity or input type.

Main Methods:

• Development of the Minerva algorithm for rule extraction.
• Application of Minerva to various black-box models.
• Benchmarking Minerva's performance against existing rule and decision tree learners.

Main Results:

• Minerva successfully extracts rules from diverse black-box models.
• The rules extracted by Minerva are comparable in performance to those from other learning methods.
• Minerva demonstrates flexibility in handling different model types and input data.

Conclusions:

• Minerva offers a significant advancement in rule extraction, enhancing the interpretability of complex machine learning models.
• The algorithm's ability to work with any black-box model broadens the applicability of interpretable AI.
• Minerva facilitates a deeper understanding of AI decision-making across various domains.