Modular Local Classification via Cluster-Guided Feature Selection in Tabular Data
Modular Local Classification via Cluster-Guided Feature Selection in Tabular Data

Summary/Abstract: Background: Many real-world tabular datasets are heterogeneous, with distinct regions of the feature space exhibiting different feature–label relationships. Conventional global classifiers often miss these local patterns, reducing both predictive accuracy and interpretability. Objective: This study aims to design a modular classification framework that combines local specialization with global consistency to enhance predictive performance and interpretability in heterogeneous tabular data. Methods: The author proposes Cluster-guided local feature selection with top-2 voting and fallback (CGLFS+), which integrates unsupervised clustering, cluster-specific feature selection and lightweight local models. Final predictions combine top-2 local decisions with a global fallback classifier for robustness. The framework was evaluated on five diverse benchmark datasets using repeated stratified cross-validation. Results: CGLFS+ achieved consistent gains in accuracy and macro F1 over strong baselines, with statistically significant improvements and competitive inference times. Conclusion: CGLFS+ successfully balances local adaptation and global consistency, providing a scalable and interpretable approach well suited to heterogeneous domains such as healthcare, chemistry and finance.

• Details
• Contents