Model-Based Learning Project

This project implements a custom Expectation-Maximization (EM) R package for clustering using Gaussian Mixture Models (GMMs). It supports flexible covariance structures and is evaluated on both low- and high-dimensional datasets, including Iris and MNIST.

---

📌 Objectives

• Implement GMM clustering with full and homoscedastic covariance models\
• Apply the EM algorithm with multiple initializations and convergence tracking\
• Use Bayesian Information Criterion (BIC) for model selection\
• Evaluate clustering performance on Iris and MNIST datasets using confusion matrices and accuracy

---

📁 Project Structure

Path	Description
MyGMMPackage_0.1.0.tar.gz	Main R package implementing the EM algorithm
R/	All R scripts for loading data and running EM
data/	Input datasets (Iris, MNIST)
results/	Output plots, confusion matrices, etc.
docs/	Project report and description (PDF format)
renv.lock	Reproducible environment snapshot
requirements.txt	List of required R packages

---

📊 Dataset Summary

Dataset	Size	Features	Classes
Iris	150	4	3
MNIST	5,000+	784 → 3/50 (after t-SNE/PCA)	10

MNIST dimensionality was reduced using PCA and t-SNE for memory size issues.

---

🧠 GMM Model Variants

• basic: Full covariance matrix per cluster\
• homoscedastic: Shared covariance matrix across clusters\
• Also supports experimental variants: CEM and SEM

---

🧪 Running the Code

1. Clone the repo

   git clone https://github.com/MedAliAdlouni/model_based_learning_project.git
   cd model_based_learning_project

2. Install required R packages

pkgs <- readLines("requirements.txt")
install.packages(pkgs)

3. Run EM on Iris dataset

source("run_EM_algorithm_IRIS_dataset.R")

4. Run EM on MNIST dataset

source("run_EM_algorithm_MNIST_dataset.R")

5. Open project report

Open report_ELADLOUNI_MOHAMMEDALI.pdf for detailed methodology, results, and analysis.