defectprediction

Defect prediction is a task in software engineering that aims to identify software modules that are likely to contain defects. The goal is to guide testing, quality assurance, and risk management by focusing effort where defects are more probable. Predictions are typically made at the module, file, or class level and can cover defects found in past releases as well as potential defects in upcoming iterations.

Data sources include bug tracking systems, version control histories, and build or test data. Features commonly used are code size metrics, complexity, churn, coupling, and historical defect counts. The labels indicate whether a defect occurred in a module within a specified window or the defect density of that module.

Approaches include traditional statistical methods such as logistic regression and more data-driven techniques such as decision trees, random forests, gradient boosting, and neural networks. Models are evaluated with metrics such as precision, recall, F1 score, and area under the ROC curve, and often require cost-sensitive evaluation to reflect testing costs. Validation should be time-aware to avoid leakage from future data.

Challenges include imbalanced defect data, noisy labels, data quality issues, and changes in development processes over time, which can cause concept drift. Defect prediction supports test planning, resource allocation, and risk assessment, enabling teams to prioritize high-risk modules for review. Common datasets used for benchmarking include the PROMISE repository and NASA’s defect datasets. Limitations include dependence on historical patterns and the risk of overfitting to past processes.