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

Area of Science:

• Software Engineering
• Machine Learning
• Data Science

Background:

• Accurate bug fix time prediction is vital for effective software maintenance and planning in open source projects.
• Evaluating various time series forecasting models is essential to enhance prediction accuracy.

Purpose of the Study:

• To compare the effectiveness of local and global time series forecasting models for predicting bug fix times.
• To assess the performance of classical models versus deep learning approaches, including global extensions with project embeddings.

Main Methods:

• Applied classical models (Naive, Linear Regression, Random Forest) and neural networks (MLP, LSTM, GRU) to real-world data from multiple repositories.
• Implemented local (per-project) and global (cross-project) modeling strategies.
• Utilized explainable AI techniques (permutation importance, saliency maps, embedding analysis) for forecast interpretation.

Main Results:

• Locally, Random Forest sometimes outperformed deep learning models in error reduction and classification metrics.
• Global models demonstrated superior robustness and generalizability.
• Global Random Forest significantly reduced mean error and maintained high accuracy and F1 score.
• Global LSTM effectively captured temporal dependencies and revealed cross-project dynamics.

Conclusions:

• An integrated, global approach combining classical and deep learning models provides more reliable and interpretable bug fix time forecasts.
• Global models enhance software maintenance planning and support by offering robust predictions.
• Explainable AI techniques are crucial for understanding forecast drivers and validating model performance.