Principal Component Analysis for Dimensionality Reduction of the Breast Cancer Dataset

Abstract

Breast cancer is a prevalent global health concern among women. This study systematically investigates the application of Principal Component Analysis for dimensionality reduction on the Wisconsin Breast Cancer Dataset. We evaluate the impact of varying PCA dimensions on the performance of several machine learning and deep learning models, including Support Vector Machine, K-Nearest Neighbors, Random Forest, Multilayer Perceptron, Fully Connected Neural Network, and Dropout models. Our findings demonstrate that PCA can enhance model performance and accuracy when dimensions are reduced from high to moderate levels. Conversely, overly aggressive dimensionality reduction leads to a significant degradation in performance (k<5). ML models exhibited optimal performance at different values of k. In which SVM achieves its best results at k=25, KNN shows stability in the range of k= 10-15, and Random Forest performs effectively at k=5. For DL models, accuracy remained stable for k ≥ 10 and saturated between k = 12-18, consistently achieving over 97% accuracy. These results underscore the critical importance of selecting an appropriate number of PCA dimensions to balance accuracy and computational efficiency, thereby improving the efficacy of breast cancer diagnostic support systems.