Data Science for the Linear Algebraist

A practical, linear-algebra-first introduction to data science.

This repository demonstrates how core linear algebra concepts -- least squares, matrix decompositions, and spectral methods -- directly power modern data science and machine learning workflows. We finish off with a mini-project involving image denoising using the truncated SVD.

Rather than treating data science as a collection of tools, this project builds everything from first principles and connects theory to implementation through Jupyter notebooks.

The compiled notebooks in this project can be viewed as a single webpage on my website. Note that if you view the notebooks in GitLab/GitHub, they have a tendency to not render the LaTeX properly.

Structure

This project is organized as a collection of focused notebooks:

images/           # saved images/visualizations
notebooks/        # jupyter notebooks containing theory, code, visuals
bibliography.md   # references for essentially everything
requirements.txt  # python requirements
LICENSE           # project license

Each notebook is self-contained and moves from theory to implementation to visualization.

Dependencies

• Python 3
• NumPy -- linear algebra
• Pandas -- data handling
• Matplotlib -- visualization
• Pillow -- imaging library
• scikit-learn -- machine learning utilities
• scikit-image -- image quality metrics

How to Run

git clone https://gitlab.com/psark/ds-for-la.git
cd ds-for-la

pip install -r requirements.txt

jupyter notebook

Open any notebook inside the notebooks/ folder.

---

Topics

1. Least Squares Regression

• Overdetermined systems
• Normal equations
• Geometric interpretation (projection onto column space)
• Implementation using NumPy

2. QR Decomposition & SVD

• Numerical stability vs. normal equations
• Orthogonal bases and conditioning
• Solving linear systems without forming X^T X

3. Some Notes & What Can Go Wrong

• Other vector norms (L^1, L^\infty), as well as matrix norms (Frobenius, Operator)
• What can go wrong?

4. Principal Component Analysis (PCA)

• Dimensionality reduction via spectral methods
• Relationship between covariance matrices and eigenvectors
• Handling correlated features

5. Project: Spectral Image Denoising via Truncated SVD

• Low-rank approximation of images
• Noise removal using singular value truncation
• RGB images (channel-wise SVD)
• Quantitative evaluation (MSE, PSNR)

6. Modelling 101

• Train/test splits
• Regression (Linear, Ridge, LASSO, PCR)
• Gradient descent
• Decision trees & random forests
• Logistic regression
• Cross-validation
• Feature scaling
• Hyperparameter tuning

---

Example: Image Denoising via SVD

Given an image matrix A (for simplicity, let's go with greyscale), we compute its singular value decomposition:

A = U \Sigma V^T

We approximate the image using only the top k singular values:

A_k = U_k \Sigma_k V_k^T

This produces:

• Noise reduction
• Compression
• A direct application of the Eckart–Young–Mirsky theorem

For color images, this is applied independently to each channel (R, G, B).

---

Key Takeaways

• Data science problems can be framed as:

  approximate solutions to linear systems

• Numerical linear algebra is necessary; it determines:

  • stability
  • performance
  • model reliability

• Spectral methods (SVD, PCA) provide:

  • structure
  • compression
  • interpretability

• Regularization connects directly to linear algebra:

  • Ridge shifts singular values, improving condition number
  • Lasso exploits L^1 geometry to produce sparse solutions

• Gradient descent convergence is governed by singular value structure

  • Condition number determines learning rate stability
  • Feature scaling reshapes the optimization landscape

---

Purpose

This project is part of a broader effort to translate a background in pure mathematics into practical data science and machine learning skills.

Future Work

• Add regularization (Ridge, LASSO)
• Extend PCA to real datasets
• Compare SVD vs. autoencoders for compression
• Add performance benchmarks (QR vs SVD vs normal equations)
• Add neural networks

---

License

This project is licensed under the MIT License. See the LICENSE file for details.