OPTIMIZING SUPERVISED MODELS FOR BENIGN-MALIGNANT BREAST TUMOR CLASSIFICATION USING ADVANCED FEATURE ENGINEERING

Abstract

Breast cancer remains a leading cause of mortality among women, necessitating advanced diagnostic tools to enhance early detection. Supervised machine learning offers promising capabilities for accurate tumor classification. This study aims to develop a robust pipeline for classifying breast tumors as benign or malignant using the Breast Cancer Wisconsin Dataset, optimizing diagnostic accuracy through advanced feature engineering. The methodology includes data preprocessing with mean imputation and min-max normalization to ensure data consistency, followed by Exploratory Data Analysis using violin plots to visualize feature distributions and correlation heatmaps to detect multicollinearity. Feature selection employed correlation-based filtering, Chi-square tests, tree-based methods, Recursive Feature Elimination (RFE), RFECV, and L1 regularization, selecting 16 key features to enhance model performance. Seven supervised algorithms—Logistic Regression, Random Forest, SVM, Extra Trees, KNN, Decision Trees, and Naïve Bayes—were trained and optimized via grid and randomized search, evaluated with stratified 10-fold cross-validation for robust assessment. Logistic Regression with Random Forest-based feature selection achieved a test accuracy of 0.977, ROC-AUC of 0.971, and average precision of 0.99, reflecting superior performance. However, L1 regularization underperformed due to over-penalization, and SVM exhibited sensitivity to feature scaling, requiring careful tuning. The proposed pipeline significantly improves diagnostic accuracy, offering a scalable, interpretable solution for clinical application.