Fuzzy Pattern Trees with Pre-classification

Fuzzy Pattern Trees evolved using Grammatical Evolution have been shown to be a robust Explainable Artificial Intelligence technique. We trained a black-box classifier, XGBoost, to identify and remove instances our evolved Fuzzy Pattern Tree model was seen to struggle to classify correctly. This framework mimics a human-computer interactive approach, where the removed instances would be classified by the human and not the Fuzzy Pattern Tree model. We investigated a range of fitness functions to ascertain which is most suitable for use with pre-classification by examining their accuracy, errors and how well they perform with the pre-classifier. We show that Fuzzy Pattern Tree classifiers, on each benchmark and using every fitness function, improved when used with a pre-classification method. Fuzzy Pattern Tree models with pre-classification found better performance than any other black-box classifier on all the benchmarks considered and routinely removed less than 10% of test data for human inspection.