Di40 / MEB-Anomaly-FW-Optimization

Implementation of three variants of the Frank-Wolfe method for solving the Minimum Enclosing Ball problem, and application to anomaly detection.

Minimum Enclosing Ball for Anomaly Detection using variants of Frank-Wolfe

Table of contents

• General info
• Team members
• Project structure
• Setup
• Usage
• Contributing

General info

This is the final project for the "Optimization for Data Science" course at the University of Padova.
This repository contains the code, report, presentation and some extra materials.

Task

1. Analyze in-depth four papers and their theoretical results.
2. Implement the Away-Steps Frank-Wolfe algorithm from Lacoste-Julien et al., 2015.
3. Implement the Blended Pairwise Conditional Gradients (BPCG) algorithm from Tsuji et. al., 2022.
4. Implement the (1+ε)-approximation to MEB algorithm from Yildirim, 2008.
5. Apply the MEB approach for anomaly detection (locate new points that are out of the MEB) on two real-world datasets.

Team members:

• Dejan Dichoski
• Marija Cveevska
• Suleyman Erim

We utilized PyCharm for development and GitHub for version control.

Project structure

├───configs
├───datasets
├───documents
│   ├───Extra papers
│   ├───Extra theory
│   ├───Papers with notes
│   └───Report and Presentation
└───src
│   ├── data_generation.py
│   ├── execution.py
│   ├── FrankWolfeVariants.py
│   ├── logger.py
│   ├── plotting.py
│   ├── utils.py
├── main.py
├── setup.py
└── requirements.txt

Setup

$ conda update conda -y
$ conda create -p venv python==3.10 -y
$ conda activate venv/

Go to the requirements.txt file and uncomment the symbol (#) before "-e .". Then:

$ pip install -r requirements.txt

Usage

Choose a configuration file (.yaml) to run an experiment from the configs folder. For example:

$ python main.py --cfg "exp6_CustomerChurn.yaml"

The results will be saved to the runs/experiment folder.

Contributing

We invite you to explore our analysis, learn from our findings, and adapt our methodologies to your own projects. Your feedback and contributions are highly valued!