Year: 2025
Type: preprint
Abstract: Rule-based models are popular for their transparency and interpretability. However, statically generated rules are often unable to accommodate changing data distributions. This paper presents a semi-supervised pipeline for automatic, adaptive rule induction via unsupervised clustering and soft label propagation, under the assumption that the data can be meaningfully clustered. If this assumption does not hold, the pipeline fails to produce reliable rules. Initially, clusters are formed from labeled data and each is assigned a human-interpretable label based on class composition. As new data exhibiting drift arrive, e.g., in batches, additional clusters are induced, and soft labels are propagated on the basis of their similarity to previously labeled clusters. A decision tree is then trained on the resulting cluster-labeled data to induce symbolic, interpretable rules, forming a surrogate model that evolves alongside the data. We evaluated three soft-label propagation strategies, proximity-based, mahalanobis distancebased, and inter/intra-cluster spread ratio within this pipeline. Rather than comparing these methods in isolation, we demonstrate how the propagation strategy significantly affects downstream rule induction. Synthetic data sets with controlled drift and publicly available datasets (real-world and synthetic) are used to assess the viability of the pipeline. Our results show that while auto-evolving rule induction is feasible and effective, its success relies on important factors like reliable label propagation strategy and the ability to keep clusterable data representations over time.
Author: Saksham Ghimire
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414080421