Predicting Software Maintenance Effort Using Historical Repository Data and Machine Learning

Abstract

Predicting software maintenance effort is an important challenge in modern software engineering because software systems continuously evolve after deployment through activities such as bug fixing, feature enhancement, and code refactoring. Software repositories store extensive historical development information including commit history, code modifications, bug reports, and issue resolution records. These repository datasets provide valuable indicators that can be analyzed to estimate maintenance workload. This study proposes a machine learning framework for predicting software maintenance effort using development activity indicators extracted from historical repository data. Features such as commit frequency, number of modified files, bug report count, developer workload, and code change size are used to train predictive models including Random Forest, Support Vector Machines, Gradient Boosting, and Neural Networks. The results show that machine learning models can effectively capture relationships between repository activity metrics and maintenance workload, enabling accurate effort prediction. The proposed framework demonstrates how repository mining combined with machine learning can support data-driven decision making and improve resource planning in software maintenance management.