Skip to content

Latest commit

 

History

History
132 lines (93 loc) · 5.28 KB

README.md

File metadata and controls

132 lines (93 loc) · 5.28 KB

This repository demonstrates how to use machine learning models to predict breast cancer diagnosis using the Breast Cancer Wisconsin (Diagnostic) Dataset in Python. It includes data preprocessing, feature engineering, model training, and evaluation.

Table of Contents

Dataset

The Breast Cancer Wisconsin (Diagnostic) Dataset is sourced from the UCI Machine Learning Repository. It contains 569 data points with 30 features, including:

  • ID: Unique identifier
  • Diagnosis: Binary label indicating diagnosis (M = malignant, B = benign)
  • radius_mean, texture_mean, perimeter_mean, area_mean, etc.: Various measurements of cell nuclei characteristics

Citation of the Dataset

Wolberg,William, Mangasarian,Olvi, Street,Nick, and Street,W.. (1995). Breast Cancer Wisconsin (Diagnostic). UCI Machine Learning Repository. https://doi.org/10.24432/C5DW2B.

Project Structure

Breast_Cancer_Diagnosis_ML/

Breast_Cancer_Diagnosis_ML/
├── data/
│   ├── breast_cancer_wisconsin.csv  # Raw dataset
├── notebooks/
│   ├── data_preprocessing.ipynb     # Data preprocessing and EDA
│   ├── feature_engineering.ipynb    # Feature engineering
│   ├── model_training.ipynb         # Model training and tuning
│   ├── model_evaluation.ipynb       # Model evaluation
├── scripts/
│   ├── preprocess.py                # Data preprocessing script
│   ├── features.py                  # Feature selection script
│   ├── train.py                     # Model training script
│   ├── evaluate.py                  # Model evaluation script
├── models/
│   ├── model.pkl                    # Trained model
├── results/
│   ├── evaluation_metrics.csv       # Evaluation metrics
│   ├── confusion_matrix.png         # Confusion matrix
├── README.md                        # Project README

Usage

  1. Data Preprocessing:

  2. Feature Engineering:

    • Run the feature_engineering.ipynb notebook to create new features and select the most important ones.
    • Feature Selection Notebook

  3. Model Training:

    • Use the model_training.ipynb notebook to train various machine learning models and tune hyperparameters.
    • Model Training Notebook

  4. Model Evaluation:

Modeling

The project explores various machine learning models, including:

  • Random Forest (RF)
  • Gradient Boosting (GB)
  • Partial Least Squares (PLS)
  • Neural Networks (NN)

The models are evaluated using cross-validation and tuned for optimal performance.

Evaluation

Model performance is assessed using metrics such as:

  • Accuracy
  • Precision
  • Recall
  • F1-Score
  • ROC-AUC

Visualizations include confusion matrices and ROC curves.

Results

Confusion Matrices

Confusion Matrices

ROC Curves

ROC Curves

Evaluation Metrics

Model Accuracy Precision Recall F1 Score ROC AUC Confusion Matrix
Random Forest 0.964912 0.953488 0.953488 0.953488 0.997707 [[69, 2], [2, 41]]
Gradient Boosting 0.947368 0.930233 0.930233 0.930233 0.995414 [[68, 3], [3, 40]]
PLS Regression 0.973684 0.976190 0.953488 0.964706 0.994759 [[70, 1], [2, 41]]
Neural Network 0.991228 1.000000 0.976744 0.988235 0.988863 [[71, 0], [1, 42]]

Discussion and Conclusion

Based on the evaluation metrics and visualizations:

  • The Neural Network model achieved the highest accuracy, precision, and F1 score, making it the best performing model in terms of overall performance.
  • Random Forest and PLS Regression also performed very well, with high accuracy and balanced precision and recall scores.
  • Gradient Boosting had the lowest accuracy and precision among the models, but still performed relatively well.
  • PLS Regression requires considerably less tuning and computing resources compared to other models, making it a suitable choice for quick and efficient predictions.
  • All models demonstrated high ROC AUC scores, indicating strong discriminative ability.

The results suggest that while the Neural Network model is the most accurate, the PLS Regression model is also a viable option when computational resources and tuning efforts are limited.

Contributing

Contributions are welcome! Please create an issue or submit a pull request for any feature requests or improvements.

License

This project is licensed under the MIT License.

If you use this repository in your research, please cite it as shown in the right sidebar.