Benchmark-Based Evaluation of Supervised Learning Models for Multi-Dataset Classification

Abstract

The performance of supervised learning models can vary considerably across datasets because of differences in dimensionality, class structure, sample size, and underlying data distributions. This study presents a controlled benchmark-based evaluation of five widely used supervised learning models: logistic regression, k-nearest neighbors, support vector machine, random forest, and gradient boosting. Four publicly available classification datasets— I​ris, Wine, Breast Cancer Wisconsin, and Digits—are used to represent low-dimensional, moderate-dimensional, binary, and multiclass prediction tasks. To ensure consistent evaluation, all models are assessed using identical stratified five-fold cross-validation partitions. Feature standardization is applied within the validation pipeline for logistic regression, k-nearest neighbors, and support vector machine, while tree-based models are trained using their original feature scales. Model performance is evaluated using accuracy, macro-precision, macro-recall, macro-F1, and ROC-AUC. The experimental results show that support vector machine achieves the highest mean macro-F1 and accuracy across the selected benchmark datasets, while k-nearest neighbors, logistic regression, and random forest remain closely competitive. Gradient boosting records comparatively lower average F1 under the selected configuration. Logistic regression achieves the highest mean ROC-AUC, demonstrating that model rankings can vary according to the selected evaluation criterion. The findings indicate that predictive model selection should be based on controlled empirical validation across multiple performance measures rather than assumptions regarding universal algorithm superiority. The study provides a reproducible framework for comparative classification experiments and supports systematic model selection across heterogeneous predictive tasks.