t-SNE Visualization on MNIST Dataset

---

🚀 Live Results

You can view the notebook with all outputs and results on Kaggle:
https://www.kaggle.com/code/evangelosgakias/t-sne-mnist-dataset

---

📑 Table of Contents

• Live Results
• Table of Contents
• Overview
• Project Structure
• Features
• Quickstart
• Usage
• Results
• Limitations and Future Work
• Contributing
• License
• Contact

---

📝 Overview

This project demonstrates the application of t-Distributed Stochastic Neighbor Embedding (t-SNE) for visualizing high-dimensional data, using the MNIST handwritten digits dataset. The notebook covers the full data science workflow: data loading, preprocessing, dimensionality reduction, visualization, and interpretation of results.

Goal:
Project high-dimensional MNIST data into 2D space to reveal clusters and patterns, and to compare t-SNE with other dimensionality reduction techniques.

---

🏗️ Project Structure

t-Distributed Stochastic Neighbor Embedding/
├── t_sne_project.ipynb      # Jupyter notebook with the complete implementation
├── requirements.txt         # Python dependencies
├── README.md                # Project documentation (this file)
├── figures/                 # Result images and plots
│   ├── sample_digits.png
│   ├── class_distribution.png
│   ├── tsne_embedding.png
│   ├── tsne_perplexity_5.png
│   └── tsne_different_perplexities.png

---

🚦 Features

Data Preparation

• Dataset Loading: Uses the MNIST dataset (70,000 samples, 784 features).
• Exploratory Data Analysis (EDA): Sample digit visualization, class distribution.
• Preprocessing: Standard scaling, subsetting for computational efficiency.

Dimensionality Reduction & Visualization

• t-SNE: 2D embedding of MNIST data, parameter tuning (perplexity).
• Visualization: Scatter plots of t-SNE embeddings, color-coded by digit label.
• Parameter Exploration: Visual comparison of t-SNE results for different perplexity values.

Interpretation & Discussion

• Cluster Analysis: Discussion of cluster structure, separation, and intra-class variation.
• Comparison: Theoretical discussion of t-SNE vs. PCA.

---

⚡ Quickstart

1. Kaggle (Recommended for Reproducibility):
   • Run the notebook on Kaggle
2. Local:
   • Clone the repo and run t_sne_project.ipynb in Jupyter after installing requirements.

---

💻 Usage

1. 📥 Clone the repository:

   git clone https://github.com/EvanGks/t-sne-mnist-visualization.git
   cd t-sne-mnist-visualization

2. 🔒 Create and activate a virtual environment:
   • Windows:

     python -m venv .venv
     .venv\Scripts\activate

   • macOS/Linux:

     python3 -m venv .venv
     source .venv/bin/activate

3. 📦 Install dependencies:

   pip install -r requirements.txt

4. 🚀 Launch Jupyter Notebook:

   jupyter notebook t_sne_project.ipynb

5. ▶️ Run all cells to reproduce the analysis and results.

🛠️ Troubleshooting:

• If you encounter missing package errors, ensure your Python environment is activated and up to date.
• For best reproducibility, use the provided Kaggle link.

---

📊 Results

Visualizations

• Sample Digits:
  
• Class Distribution:
  
• t-SNE Embedding (Perplexity=30):
  
• t-SNE with Different Perplexities:
  

Key Findings

• t-SNE reveals clear clusters corresponding to digit classes in the MNIST dataset.
• The choice of perplexity significantly affects the visualization; values between 30 and 50 often yield the best separation for MNIST.
• Some digits (e.g., 4 and 9, 3 and 8) may have overlapping clusters due to visual similarity.
• t-SNE excels at local structure but does not always preserve global distances.

For full details, metrics, and plots, see the notebook.

---

📝 Limitations and Future Work

• Computational Cost: t-SNE is slow for large datasets; subsetting is used for efficiency.
• Parameter Sensitivity: Results depend on perplexity and other parameters.
• Global Structure: t-SNE is best for local structure; global distances may be misleading.
• Potential Improvements:
  • Compare with UMAP and PCA visually and quantitatively.
  • Automate parameter selection.
  • Apply to other high-dimensional datasets.

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request or open an Issue for suggestions and improvements.

---

📝 License

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

---

📬 Contact

For questions or feedback, please reach out via:

• GitHub: EvanGks
• X (Twitter): @Evan6471133782
• LinkedIn: Evangelos Gakias
• Kaggle: evangelosgakias
• Email: evangks88@gmail.com

---

Happy Visualizing!