An Exploration of FOLD-R++: Using Normal Logic Programming to Create Explainable AI Models

An Exploration of FOLD-R++: Using Normal Logic Programming to Create Explainable AI Models

This paper explores FOLD-R++, an inductive logic programming algorithm that learns default rules with exceptions — a structure that mirrors human reasoning more closely than typical black-box classifiers — and generates output as readable logic programs rather than opaque weights. Paired with s(CASP), a constraint answer-set solver, the system can also produce step-by-step justifications and proof trees explaining why a specific prediction was made, breaking each rule down into which conditions held true or false for that instance.

The method was tested on two classification tasks: predicting loan defaults (Kaggle’s Loan Default dataset) and predicting Titanic passenger survival. After tuning the algorithm’s fit ratio, the loan-default model reached 88.2% accuracy and the Titanic model reached 80.5% — both while remaining fully interpretable, with every prediction traceable to a specific, human-readable rule chain. On the loan-default task, this meant a bank could see exactly which applicant attributes (interest rate, credit score, loan-to-value ratio, etc.) drove a decision; on the Titanic task, the resulting rules surfaced known socioeconomic patterns in survival, such as fare and class correlating with outcome, while still correctly identifying exceptions to that trend rather than discarding them.

Written during a research internship in UT Dallas’s Applied-Logic, Programming Languages & Systems group, under Dr. Gopal Gupta (Head, Department of Computer Science, UT Dallas), Summer 2023. Related: an accompanying code walkthrough is available on Kaggle.