Feature Selection Benchmarks for Breast Cancer Diagnosis: A Comparative Machine Learning Study

Abstract

Breast cancer remains one of the most common causes of death among women, making early and precise detection essential. Yet conventional diagnosis can be limited by specialist shortages, cost, and slow workflows. We therefore assess machine-learning classification with feature selection to streamline diagnosis. Our contribution is a comparative benchmark of feature-selection strategies and classifiers on the WDBC dataset. We evaluated five models (SVM, neural-networks, decision tree, bagged-tree, and boosted-tree). Chi², mRMR, and ReliefF selected 5, 10, 15, and 30 features, and performance was measured across multiple train–test splits using accuracy, precision, recall, specificity, and F1-score. SVM was overall the top performer and stable across splits. The best SVM setting reached 97.81% accuracy, with strong precision and F1-score, indicating reliable benign–malignant separation. Neural-networks usually ranked second but were more sensitive to the split. Bagged trees generally improved on a single decision tree, while boosted trees showed mixed gains depending on the subset. ReliefF and mRMR often matched or exceeded Chi² with smaller subsets, showing that careful feature reduction can retain accuracy while lowering dimensionality. In conclusion, combining effective feature selection with an appropriate classifier improves breast cancer classification, and SVM with a compact feature set is a practical choice.