Linear SVM from Scratch (Portfolio Sample)

This repository implements a soft-margin linear SVM from scratch using NumPy, trained by batch gradient descent on the hinge loss with L2 regularization. It also includes a brief Gaussian kernel Gram matrix demo (educational) and clean plots of the learned decision boundary on synthetic data.

What this shows

• From-scratch optimization of hinge loss + L2 (soft-margin SVM)
• Clear, reproducible notebook with decision boundary visualizations
• Educational Gram matrix (RBF) construction
• Comparison to scikit-learn's LinearSVC as a sanity check

Structure

.
├── notebooks
│   └── svm_from_scratch.ipynb   # Training, plots, Gram matrix demo
├── src
│   └── svm_scratch.py           # Minimal SVM class (NumPy)
├── README.md
├── requirements.txt
├── LICENSE
└── .gitignore

Quickstart

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/Mac: source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook notebooks/svm_from_scratch.ipynb

Notes

• The scratch model is linear; the RBF Gram matrix cell is for intuition/visuals, not a full kernel-SVM solver.
• Plots are made with matplotlib only.
• Code is kept concise and readable for portfolio review.

License

MIT