<![CDATA[Although dimensionality reduction is widely incorporated into machine-learning pipelines for breast cancer diagnosis, its independent contribution remains unclear when classifiers are optimized under identical experimental conditions. This study evaluates Support Vector Machine (SVM) and Principal Component Analysis–Support Vector Machine (PCA-SVM), both optimized through Random Search, to determine whether PCA provides a measurable classification benefit. The Wisconsin Diagnostic Breast Cancer dataset, consisting of 569 samples, 30 numerical features, and benign and malignant classes, was analyzed. Preprocessing included Interquartile Range-based outlier removal, label encoding, Min-Max normalization, and Synthetic Minority Over-sampling Technique applied to the training data. PCA retained 95.30% of cumulative explained variance and reduced the feature space from 30 variables to 10 principal components. Both pipelines used identical preprocessing, optimization, evaluation metrics, and train-test partitions of 60:40, 70:30, and 80:20. Performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. The optimized SVM achieved its best performance at the 70:30 split, with 98.33% accuracy, 98% precision, 98% recall, 98% F1-score, and a ROC-AUC of 0.9998. Under the same split, PCA-SVM reached 97.22% accuracy and remained below SVM across the reported metrics. Despite reducing dimensionality by 66.7%, PCA did not improve predictive performance, indicating that original features retained important class-discriminative information. These findings show that preserving the original feature representation was more effective than PCA-based reduction and that dimensionality reduction should be justified empirically rather than adopted as a routine preprocessing step.]]>
