2025
Breast cancer diagnosis with DNNs, SVMs and PCA
Three approaches on Breast Cancer Wisconsin, with a Keras neural network, a cross-validated comparison of regularized models, and dimensionality reduction with PCA.
Problem
Classify tumors as benign or malignant from cell measurements.
Data
Two versions of the Wisconsin breast cancer data. The cross-validation study uses *Diagnostic* (569 tumors, 30 features, bundled with scikit-learn). The DNN and the PCA use a course-provided version of *Original* (583 samples, 9 continuous cell features, 65% benign), which is not included in the repo.
Method
A DNN with a search over layers, units, dropout and optimizer; L1/L2 logistic regression, SVMs and trees with cross-validation; PCA with a scree plot and a 90%-variance criterion.
Result
The L2-regularized SVM reached 97.6% cross-validated accuracy and 94.7% on test (F1 0.947), the DNN 94.5% on test, and 3 principal components keep 81.9% of the variance.
Stack
- Python
- TensorFlow
- Keras
- scikit-learn
- pandas
Context
Course: Machine Learning II · Data Mining · 2025