Identifying 'Candidates for University Projects' in Open Source Software

This repository contains the complete code, data, and results for reproducing the experiments described in the paper "Identifying 'Candidates for University Projects' in Open Source Software". The work presents a machine learning approach to automatically identify GitHub issues suitable for university student projects. We compare Random Forest classifiers with Large Language Model (LLM) baselines for both binary classification (CUP vs. NOT CUP) and multi-label property prediction (implementation effort, testing effort, issue understanding effort, main focus, and size of project). The repository includes: - Complete Python codebase (29 scripts) for data processing, model training, and evaluation - Pre-processed datasets: 3,295 GitHub issues from JabRef (53 labeled as Candidates for University Projects) - Trained Random Forest models for property prediction - All experimental results matching the paper's tables (Tables 1-8) - Comprehensive documentation for reproducibility Key Results: - Random Forest achieved 45% recall (5/11) at K=20 for binary classification - Maintainer review confirmed 13/25 new candidate issues identified by the model - Property prediction achieved F1-scores ranging from 0.022 (Labels) to 0.395 (Size of project) using cosine similarity - LLM baseline showed limited effectiveness (recall: 0.091-0.182) compared to supervised learning All results are fully reproducible with the provided code and data.