Hybrid Transformer Network for Predicting Software Maintenance Effort from Source Code Evolution

Abstract

Software maintenance effort is difficult to estimate because code changes, developer activity, file dependencies, and earlier maintenance patterns evolve across releases. This study presents a hybrid transformer network for predicting maintenance effort from source-code evolution. The model combines temporal repository events, structural code information, numerical evolution features, and historical maintenance time through separate learning branches. A gated fusion layer integrates these representations and produces effort estimates in developer-hours. The model was evaluated across eight successive release cycles and compared with a conventional recurrent network. Results showed that the hybrid transformer followed the actual maintenance pattern more closely, especially during releases with high code churn, many modified files, dependency changes, and developer handovers. It also reduced large prediction errors during complex maintenance periods. The proposed approach can support workload estimation, developer assignment, release planning, and maintenance budgeting. It provides a practical method for improving effort prediction accurately in evolving software projects.