Breast Cancer Classification Comparison

A comparative study of seven classic machine learning algorithms for breast cancer diagnosis on the Wisconsin Breast Cancer dataset — going beyond a single accuracy score to look at precision/recall trade-offs, cross-validated stability, ROC-AUC, and feature importance.

Overview

This project trains and evaluates seven classifiers on an identical, leakage-free preprocessing pipeline, then compares them side by side:

• Gaussian Naive Bayes
• k-Nearest Neighbors (kNN)
• Decision Tree
• Random Forest
• Support Vector Machine (SVM)
• Logistic Regression
• Artificial Neural Network (MLP)

In a diagnostic setting, missing a malignant case (low recall) is far more costly than a false alarm — so the comparison looks at precision and recall separately, not just overall accuracy.

Dataset

The Breast Cancer Wisconsin (Diagnostic) dataset, loaded directly via sklearn.datasets.load_breast_cancer():

• 569 samples, 30 numeric features (radius, texture, perimeter, area, smoothness, etc., computed from digitized images of fine needle aspirates)
• Binary target: malignant / benign

As the PCA projection above shows, the two classes separate fairly cleanly even in just two principal components — a good early signal that most classifiers should perform well here.

Pipeline

1. Explore the dataset: class balance, and correlation between the highest-variance features.
2. Split into train (80%) / test (20%) with train_test_split, using a fixed random_state and stratify to keep results reproducible and class-balanced.
3. Scale features with MinMaxScaler, fit only on the training set and applied to the test set — no data leakage.
4. Train all seven classifiers on the same scaled training data.
5. Evaluate each model on training accuracy, test accuracy, precision, recall, and F1-score.
6. Validate more rigorously with 5-fold stratified cross-validation, so results aren't dependent on one lucky/unlucky split.
7. Visualize everything: metric comparisons, a radar chart, confusion matrices, ROC curves, feature importance, and learning curves.

Results

Model	Train Acc.	Test Acc.	Precision	Recall	F1-score
ANN	0.989	0.983	0.986	0.986	0.986
SVM	0.987	0.974	0.986	0.972	0.979
Logistic Regression	0.978	0.956	0.947	0.986	0.966
Random Forest	1.000	0.956	0.959	0.972	0.966
KNN	0.980	0.956	0.972	0.958	0.965
Naive Bayes	0.939	0.930	0.944	0.944	0.944
Decision Tree	1.000	0.912	0.956	0.903	0.929

(Single 80/20 split, random_state=42. See the cross-validation section below for a more robust comparison.)

Precision, Recall & F1-score

Multi-metric snapshot

Confusion matrices

ROC curves

Feature correlation

Random Forest and Decision Tree perfectly fit the training data (1.000), which is a classic overfitting signature — their lower test scores compared to ANN and SVM reflect that. ANN and SVM generalize the best on this split, though the cross-validation results in the notebook show all models are fairly close and stable overall.

Requirements

numpy
pandas
scikit-learn
matplotlib
seaborn

Install with:

pip install numpy pandas scikit-learn matplotlib seaborn

Usage

git clone https://github.com/sadrayef/breast-cancer-classification-comparison.git
cd breast-cancer-classification-comparison
jupyter notebook Breast_cancer.ipynb

Run all cells in order. The train/test split and all stochastic models use a fixed random_state, so results are reproducible across runs.

Notes & Possible Extensions

• Hyperparameters are set manually rather than tuned via GridSearchCV or RandomizedSearchCV.
• The dataset is fairly small and clean; results might shift with a noisier, real-world clinical dataset.
• Model explainability (e.g. SHAP values) would be a natural next step beyond basic feature importance.

Disclaimer

This project is for educational and research purposes as part of a machine learning coursework assignment. It is not intended for clinical or diagnostic use.

License

This project is open source and available under the MIT License.