Learning Interpretable Sequence Classifiers through Evolved Regular Expression Patterns

Abstract

Sequence classification methods based on deep neural networks achieve strong predictive performance but often lack interpretability. This limitation is critical in domains where decision transparency is required. This work investigates a preliminary framework for supervised classification of symbolic sequences in which models are expressed directly as Regular Expressions (Regex). Classification is formulated as a pattern-matching problem, and candidate symbolic rules are discovered through a search-based induction process. Preliminary experiments on controlled synthetic datasets suggest that evolving symbolic expressions can capture patterns consistent with the underlying generative structure while producing inherently interpretable classifiers. These findings highlight the potential of symbolic pattern evolution for transparent sequence learning.