PCA-Based Digit Visualization

This repository demonstrates how Principal Component Analysis (PCA) can be applied to pixel-based handwritten digit images.
The project focuses on dimensionality reduction, visual interpretation of components, and understanding how PCA transforms high-dimensional image data into a meaningful latent space.

---

Overview

This project applies Principal Component Analysis (PCA) to the classic Handwritten Digits dataset.
The workflow focuses on:

• Extracting pixel data from an 8×8 grayscale image grid
• Visualizing sample digits as heatmaps and images
• Reducing 64-dimensional pixel data to 2D and 3D PCA space
• Understanding variance explained by principal components
• Visualizing digit clusters based on PCA compressed features

PCA is especially powerful in high-dimensional image tasks where each pixel is a feature.

---

Dataset

• Name: Digits Dataset (Scikit-Learn)
• Source: Data set
• Records: 1,797 handwritten digits (0–9)
• Features:
  • 64 pixel intensity values (8×8 image)
  • number_label (target digit)

Each image is stored as a flattened array of 64 grayscale values.

---

Project Workflow

1) Loading the Dataset

Data is loaded from CSV containing pixel values and the digit label.

• Display dataset info
• Verify 64 pixel columns
• Confirm no missing values

2) Pixel Extraction

Created a new DataFrame:

pixels = df.drop("number_label", axis=1)

Extracted a single digit representation (first row), converted to NumPy array, and reshaped it into 8×8 grid.

3) Visualizing Digits

Three visualizations:

• matplotlib.imshow (default colormap)
• matplotlib.imshow(cmap='gray')
• seaborn.heatmap with pixel intensities

These help interpret the pixel intensities and confirm data structure.

4) Scaling Pixel Features

Used:

StandardScaler()

Scaling is essential because PCA relies on variance, and raw pixel ranges differ across images.

5) PCA with 2 Components

Performed dimensionality reduction:

• PCA(n_components=2)
• Projected all digits to 2D PCA space
• Visualized using a color-coded scatter plot (hue = digit label)

Digits form distinct natural clusters, even with only 2 components.

Explained variance of PC1+PC2:

~21.59%

6) PCA with 3 Components

Extended the model:

• PCA(n_components=3)
• 3D scatter plot using Matplotlib's 3D axis
• Color-coded by digit label

This provides an even clearer separation for some digit classes.

---

Libraries Used

• numpy
• pandas
• seaborn
• matplotlib
• scikit-learn (PCA, StandardScaler)

---

How to Run

Clone the repository:

GitHub Repository

Install dependencies:

pip install -r requirements.txt

• requirements.txt → File

or directly:

pip install numpy pandas seaborn matplotlib scikit-learn

Run the script to generate all visualizations.

---

Results Summary

PCA with 2 Components

• Captures ~21.6% of the variance.
• Despite low variance percentage, digits form recognizable clusters.
• Demonstrates PCA’s ability to compress images while retaining structure.

PCA with 3 Components

• Better separation of digits in 3D space.
• Useful for interactive visualization and cluster analysis.

Image Reconstruction Insight

Each digit's raw pixel row can be reshaped back into an 8×8 grid to visually confirm the sample.

---

Key Takeaways

• PCA can cluster handwritten digits even with limited components
• 64-dimensional pixel data compresses cleanly into 2D and 3D
• Variance captured by early components contains meaningful structure
• PCA is suitable for visualization, preprocessing, and noise reduction

---

Conclusion

This project demonstrates how PCA transforms high-dimensional pixel data into low-dimensional latent space:

• Meaningful digit clusters emerge even with 2 or 3 components
• Useful for visualization, feature extraction, and preprocessing for ML models
• Highlights the power of dimensionality reduction on image datasets

PCA remains a foundational tool in exploratory data analysis for image-based machine learning tasks.

---

Author ✍️

Author: Ali
Field: Data Science & Machine Learning Student
Email: ali.hz87980@gmail.com
GitHub: ali-119