A Comparative Analysis of Feature Selection Approaches on Base and Ensemble Machine Learning Models in Breast Cancer Diagnosis

Abstract

In this study, combinations of machine learning and ensemble models using different feature selection techniques were examined on the Breast Cancer Wisconsin dataset, and the differences in performance of these models across various feature selection techniques were compared. Logistic Regression (LR), Support Vector Machines (SVM), Multilayer Perceptron (MLP), k-Nearest Neighbor (k-NN), Decision Trees (DT), Random Forest (RF), Naive Bayes (NB), XGBoost, LightGBM and ensemble learning methods were used in the study. Two different feature selection methods (Filter and Wrapper) were applied to optimize model performance. To prevent data leakage, feature selection operations were performed only on the training data. To validate the models, 5-fold cross-validation was performed with 10 different seed values. The performance of the predictions was evaluated using Accuracy, Precision, Recall, Specificity, F1 Score, Matthews Correlation Coefficient (MCC) and ROC-AUC. The models with the highest accuracy obtained from various combinations of feature selection and machine learning were analyzed using the nonparametric Friedman test, and the difference between the two most successful models was evaluated using the Nadeau & Bengio-corrected t-test. Finally, the decision-making process of the most successful model was explained using SHAP (Shapley Additive exPlanations) analysis for transparency and interpretability.