Cross-Dataset Evaluation of Support Vector Machines: A Reproducible, Calibration-Aware Baseline for Tabular Classification

Support Vector Machines (SVMs) remain competitive for small and medium-sized tabular classification problems, yet reported results on benchmark datasets vary widely due to inconsistent preprocessing, validation, and probability calibration. This paper presents a calibration-aware, cross-dataset benchmark that evaluates SVMs against classical baselines—Logistic Regression, Decision Tree, and Random Forest—under leakage-safe pipelines and statistically principled protocols. Using three representative binary datasets (Titanic survival, Pima Indians Diabetes, and UCI Heart Disease), we standardize imputation, encoding, scaling, and nested cross-validation to ensure comparability. Performance is assessed not only on discrimination metrics (accuracy, precision, recall, F1, PR--AUC) but also on probability reliability (Brier score, Expected Calibration Error) and threshold optimization. Results show that tuned RBF--SVMs consistently outperform Logistic Regression and Decision Trees, and perform comparably to Random Forests. Calibration (Platt scaling, isotonic regression) substantially reduces error and improves decision quality, while domain-specific features enhance Titanic prediction. By embedding all steps in a transparent, reproducible protocol and validating across multiple datasets, this study establishes a rigorous methodological baseline for SVMs in tabular binary classification, providing a reference point for future machine learning research.