Time series forecasting for bug resolution using machine learning and deep learning models.

Predicting bug fix times is a key objective for improving software maintenance and supporting planning in open source projects. In this study, we evaluate the effectiveness of different time series forecasting models applied to real-world data from multiple repositories, comparing local (one model per project) and global (a single model trained across multiple projects) approaches. We considered classical models (Naive, Linear Regression, Random Forest) and neural networks (MLP, LSTM, GRU), with global extensions including Random Forest and LSTM with project embeddings. The results highlight that, at the local level, Random Forest achieves lower errors and better classification metrics than deep learning models in several cases. However, global models show greater robustness and generalizability: in particular, the global Random Forest significantly reduces the mean error and maintains high performance in terms of accuracy and F1 score, while the global LSTM captures temporal dependencies and provides additional insights into cross-project dynamics. The explainable AI techniques adopted (permutation importance, saliency maps, and embedding analysis) allow us to interpret the main drivers of forecasts, confirming the role of process variables and temporal characteristics. Overall, the study demonstrates that an integrated approach, combining classical models and deep learning in a global perspective, offers more reliable and interpretable forecasts to support software maintenance.