Analyzing productivity in open source software development: an explainable AI approach using SHAP and causal discovery

• Ano de defesa: 2025
• Autor(a) principal: Gut, Christian Manfred Toni
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Algoritmo FCI; Causal discovery; Causal inference; Descoberta causal; Empirical software engineering; Engenharia de software empírica; Explainable AI; FCI algorithm; Framework SPACE; GitHub; Inferência causal; Inteligência artificial explicável; Métricas de produtividade; Mineração de repositórios de software; Mining software repositories; Open source software; Productivity metrics; SHAP; Software de código aberto; SPACE framework
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-01092025-084702/

As software becomes increasingly important, there is a growing focus on developing it in an effective and efficient manner. Thereby Open Source software is of particular interest for researchers due to the transparency of its development process. This thesis leverages this transparency by subjecting data mined from GitHub to explainable AI methods in order to understand what factors influence the productivity in Open Source software development. To address the complexity of software development, a multidimensional approach, based on the SPACE framework, was adopted, using five different productivity metrics: Developer Churn, Stars Added, LOC per Developer, Merge Times and Issue Solutions Times. Taking advantage of the wealth of data from the Git commit history and GitHub API, a variety of input variables were generated to address the team structure, the structure of the source code, and software engineering practices. These variables served to train machine learning models to predict each of the five productivity metrics. Using SHapley Additive exPlanations (SHAP), the importance of input variables could be assessed, providing insight into which factors may be most critical for each productivity metric. To provide additional understanding, the most important variables were also used as input for the FCI algorithm, a computational method that helps to uncover causal relationships. The analysis of the SHAP values and causal graphs from the FCI algorithm concluded that the traditional understanding of measuring productivity in software development may not fully apply to Open Source, because sporadic contributions may lead to misinterpretations of productivity metrics. The thesis also identified File Age as a relevant factor, exerting a positive influence on multiple productivity metrics if new features are being developed, compared to the maintenance of existing features. Lastly, Closeness Centrality, a network metric derived from team interactions and collaborations, was found to exert a significant positive correlation or even influence on various productivity measures.