Hypergraph-Driven Semi-Supervised Learning Approaches for Variants of SVMs

This repository contains the LaTeX source and supporting files for my Master's thesis, "Hypergraph-driven Semi-Supervised Learning Approaches for Variants of SVMs", completed at the South Asian University under the supervision of Dr. Reshma Rastogi.

Thesis PDF

You can download the compiled thesis PDF here.

DOI: Research Gate

Blog post discussing project and my contribution: Medium

Project Overview

The thesis explores two novel semi-supervised frameworks:

1. Improved Hypergraph Laplacian Support Vector Machine (IHLSVM): Combines Laplacian and hypergraph representations to enhance pairwise and higher-order interactions for robust pattern classification.
2. Hypergraph Regularized Semi-Supervised Least Squares Twin Support Vector Machine (HMLLSTSVM): A multilabel learning framework leveraging hypergraph Laplacians and least-squares loss, particularly effective for sparse label scenarios.

Applications

The proposed methods excel in real-world tasks such as:

• Medical diagnosis
• Text classification
• Image annotation

Key Features

• Novel Algorithms: IHLSVM and HMLLSTSVM.
• Benchmark Evaluations: Validated on standard datasets like Emotions, Flags, Image, and Yeast.
• State-of-the-Art Comparisons: Demonstrates superior classification and multilabel learning performance.

Repository Structure

• disssertation_main.tex: LaTeX source for the thesis.
• figures/: Figures used in the thesis.
• bibliography.bib: Bibliography file for references.
• output/: Directory for compiled PDF output.

Prerequisites

To compile the LaTeX source, ensure you have:

• A LaTeX distribution (e.g., TeX Live, MiKTeX).
• A LaTeX editor (e.g., Overleaf, Texmaker).

How to Compile

1. Clone the repository:

   git clone https://github.com/devn913/master-thesis
   cd master-thesis

2. Compile the LaTeX file:

   pdflatex disssertation_main.tex
   bibtex disssertation_main
   pdflatex disssertation_main.tex
   pdflatex disssertation_main.tex

3. Alternatively, use Overleaf for online editing and compilation.

Publications

• R. Rastogi and D. Nirwal , “Hypergraph Regularized Semi-Supervised Least Squares Twin Support Vector Machine for Multilabel Classification,” Lecture Notes in Computer Science, Dec. 2024. (ICPR)
• R. Rastogi and D. Nirwal, “Improved Hypergraph Laplacian Based Semi-Supervised Support Vector Machine,” Lecture Notes in Computer Science, Dec. 2024. (ICPR)

Results

Multi-label Dataset Summary

Dataset	Instances	Features	Labels	Cardinality
Emotions	593	72	6	1.87
Flags	194	19	7	3.392
Image	2000	294	5	1.24
Yeast	2417	103	14	4.24
Birds	645	260	19	2.08
CAL500	502	68	174	26.04
Enron	1702	1001	53	3.38
Scene	2407	294	6	1.07

Performance Highlights

• IHLSVM: Outperforms existing methods for binary and multi-class classification.
• HMLLSTSVM: Achieves state-of-the-art results in multilabel learning with sparse data.

License

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

Acknowledgments

Special thanks to:

• Dr. Reshma Rastogi, my supervisor, for her invaluable guidance.
• My colleagues and labmates for their constant support.
• Laboratory: Machine Learning and Statisical Inference (MLSI)

Author

Dev Nirwal
Your GitHub Profile