Thesis Repository: "Application of Machine Learning Techniques to Improve the Performance of Graph-Based Information Retrieval Models"

This repository hosts the code implementation, experiments, and results for my Integrated Master's thesis, titled "On the Improvement of Graph-Based Information Retrieval Models Using Machine Learning", along with its byproduct—a paper titled "Spectral Clustering and Query Expansion on the Graphical Set-Based Model". For these purposes, a Python library named infre was developed, supporting classical Information Retrieval (IR) models such as:

• Vector Space (VS) [Link to VS model reference]
• Set-Based (SB) [Link to SB model reference]

Overview of the infre Library

In this study, a new family of IR models was designed and implemented by extending the SB and Graphical extension of the Set-Based (GSB) models. The infre library is envisioned to evolve, hosting various state-of-the-art (SOTA) IR models. The goal is to allow easy fitting to specific collections and seamless comparison across models using IR metrics like Precision, Recall, F1-Score, AVEp, MAP, DCG, etc., and others for educational and research purposes.

Repository Structure

• experiments/: This folder contains all the Python scripts (.py files) used for running the associated experiments. Results are exported as .xlsx files for further analysis.
• notebooks/: The final analysis is performed through Jupyter notebooks (with filenames containing *-exploitation). These notebooks load the .xlsx files, compare the results across models, and validate our hypothesis—demonstrating that introducing a conceptualized aspect into graph-based IR models improved precision by x%.

Example Usage

• main.py: Serves as an example for instantiating, running, and evaluating the models on the CF collection. It demonstrates the ease of using the infre library for benchmarking different IR models.

Future Vision

The future goal of the infre library is to become a comprehensive tool for IR research, facilitating the implementation and comparison of a wide range of IR models. As it evolves, it will aim to incorporate more SOTA methods and provide a flexible platform for testing new ideas in the field of Information Retrieval.

---

Feel free to clone, explore, and contribute to the project!