Breast Cancer Classification: A Comparative Machine Learning Study

A complete, end-to-end machine learning project that trains and evaluates six classification algorithms on the Wisconsin Breast Cancer dataset. The goal is to identify which algorithm most reliably distinguishes malignant from benign tumors using cell nucleus measurements from biopsy samples.

---

Table of Contents

• Project Overview
• Dataset
• Project Structure
• Machine Learning Pipeline
• Results
• Key Visualizations
• How to Run
• Requirements
• Limitations
• Author

---

Project Overview

Most machine learning tutorials train one model and stop. This project takes a different approach. Six algorithms are trained under identical conditions, validated through 10-fold stratified cross-validation, and the top three candidates are tuned with GridSearchCV. Every decision in the pipeline is justified, every result is traceable, and every chart is saved to disk.

The project follows the industry-standard ML workflow from raw data to a final recommendation backed by multiple evaluation methods.

---

Dataset

Source: Wisconsin Breast Cancer Dataset via sklearn.datasets

Property	Value
Samples	569 patients
Features	30 numerical features
Target classes	Malignant (0), Benign (1)
Class distribution	37.3% Malignant, 62.7% Benign
Missing values	None

Features are computed from digitized images of fine needle aspirate (FNA) biopsy samples. They describe characteristics of cell nuclei including radius, texture, perimeter, area, smoothness, compactness, concavity, symmetry, and fractal dimension. Each characteristic is represented three times: as a mean, a standard error, and a worst (largest) value.

---

Project Structure

breast-cancer-classification/
│
├── breast_cancer_classification.ipynb   # Main notebook
│
├── images/                              # All charts saved automatically
│   ├── 01_class_distribution.png
│   ├── 02_mean_feature_distributions.png
│   ├── 03_violin_plots_by_class.png
│   ├── 04_kde_by_class.png
│   ├── 05_feature_separation_scores.png
│   ├── 06_correlation_matrix.png
│   ├── 07_target_correlation.png
│   ├── 08_pair_plot_top_features.png
│   ├── 09_scaling_comparison.png
│   ├── 10_baseline_model_comparison.png
│   ├── 11_malignant_recall_comparison.png
│   ├── 12_cv_f1_distribution.png
│   ├── 13_cv_mean_scores.png
│   ├── 14_tuning_comparison.png
│   ├── 15_confusion_matrices.png
│   ├── 16_roc_curves.png
│   ├── 17_learning_curve.png
│   ├── 18_feature_importance.png
│   └── 19_final_overview.png
│
└── README.md

---

Machine Learning Pipeline

Phase	Description
1	Project setup: imports, global settings, output folder
2	Data loading and audit: shape, types, nulls, class balance
3	EDA: distributions, violin plots, KDE by class, separation scores
4	EDA: correlation matrix, target correlation, pair plots
5	Preprocessing: stratified train-test split, StandardScaler
6	Baseline comparison: six algorithms trained and ranked
7	Cross-validation: 10-fold stratified CV for all six models
8	Hyperparameter tuning: GridSearchCV on top three models
9	Final evaluation: classification reports, confusion matrices, ROC curves, learning curves
10	Conclusion: feature importance, full overview, limitations, ethical considerations

---

Results

Baseline Comparison (Test Set, No Tuning)

Model	Accuracy	F1-Score	Malignant Recall
Logistic Regression	0.9825	0.9861	0.9762
Support Vector Machine	0.9825	0.9861	0.9762
K-Nearest Neighbors	0.9561	0.9655	0.9286
Random Forest	0.9561	0.9655	0.9286
Naive Bayes	0.9298	0.9444	0.9048
Decision Tree	0.9123	0.9286	0.9286

Cross-Validation Results (10-Fold Stratified)

Model	CV F1 Mean	CV F1 Std
Logistic Regression	0.9859	0.0109
Support Vector Machine	0.9791	0.0103
K-Nearest Neighbors	0.9742	0.0115
Random Forest	0.9707	0.0261
Naive Bayes	0.9463	0.0319
Decision Tree	0.9407	0.0317

Final Evaluation - Tuned Models

Model	F1-Score	Malignant Recall	ROC-AUC	Missed Cases
Support Vector Machine	0.9861	0.9762	0.9977	1 of 42
Logistic Regression	0.9861	0.9762	0.9960	1 of 42
Random Forest	0.9655	0.9286	0.9932	3 of 42

Recommended model: Support Vector Machine (C=10, gamma=0.01, kernel=rbf)

It leads on ROC-AUC at 0.9977, matches the highest F1-Score and accuracy, and has the lowest cross-validation standard deviation of all six models at 0.0103. That combination of peak performance and consistency makes it the most reliable classifier in this study.

---

Key Visualizations

Class Distribution

Feature Separation by Class

Cross-Validation F1 Distribution

Confusion Matrices

ROC Curves

Learning Curve

---

How to Run

1. Clone the repository

git clone https://github.com/arberzylyftari/breast-cancer-classification.git
cd breast-cancer-classification

2. Install dependencies

pip install numpy pandas matplotlib seaborn scikit-learn jupyter

3. Launch the notebook

jupyter notebook breast_cancer_classification.ipynb

4. Run all cells

From the menu bar select Kernel, then Restart and Run All. The notebook runs top to bottom with no errors. All charts are saved automatically to the images/ folder.

No external dataset download is required. The Wisconsin Breast Cancer dataset is loaded directly from scikit-learn.

---

Requirements

numpy
pandas
matplotlib
seaborn
scikit-learn
jupyter

Install all at once:

pip install numpy pandas matplotlib seaborn scikit-learn jupyter

Tested with Python 3.8 and above.

---

Limitations

• The dataset comes from a single institution. Performance may differ on data from different imaging equipment or patient populations.
• The test set contains 114 samples. Confidence intervals around reported metrics are wider than the point estimates suggest.
• This model is a diagnostic aid only. It is not a replacement for clinical judgment and should never be used as the sole basis for a medical decision.
• No external validation was performed. The test set was drawn from the same distribution as the training data.

---

Author

Built as part of a machine learning portfolio at Holberton School.

For questions or feedback, reach out via GitHub.