Sinusoidal time-based features and human error metrics: Advancing software defect prediction in safety-critical systems

Defect detection in safety-critical software remains difficult despite advanced tools and mature quality assurance, largely due to the human origins of many errors. Building on prior work introducing human error–driven metrics that outperform traditional code measures, this study enhances predictive accuracy by prioritizing higher recall to strengthen defect triage in environments where missing defects carries severe risk and moderate false positives are acceptable. We integrate temporal cyclicality into defect prediction by transforming code commit timestamps into sine features via a parameterized sinusoidal model, optimized with a genetic algorithm to capture daily and periodic developer activity patterns. These features preserve non-linear, cyclical relationships linked to defect introduction, allowing machine learning models to exploit latent human-behavioural signals. Evaluation across three open-source safety-critical systems shows average recall gains of 48.68% over code metrics baselines and 9.27% over previously defined human error metrics. Embedding periodic human activity patterns alongside human-error features significantly improves defect prediction. The approach is interpretable, and generalizable, offering a pathway for broader application and future integration with adaptive, human-centric software quality models.