Explainable-ML-based Framework to Predict Software Defects in Real-Time Control Systems

Machine Learning (ML) algorithms have significantly improved software engineering activities by examining software metrics and repositories, leading to more accurate predictions and improved software quality. Here, we have used five Machine learning models, which are Random Forest, Bootstrap aggregating, Extreme Gradient Boosting algorithm and K-Nearest Neighbours algorithm, to compare which performs the best in software fault prediction, specifically targeting real-time control systems where reliability and responsiveness are critical. Comparing the results of the model we infer that the boosting algorithm has the highest accuracy, while still maintaining other evaluation scores. After thorough evaluation we got the maximum accuracy of 0.78 with the bagging algorithm. We have also added a layer of transparency by using X-AI. For this, we have used Shapley Additive Explanation (SHAP) and Local Interpretable Model (LIME). The results from both techniques support each other, and the feature which has the maximum impact is the number of lines of code. X-AI enhances predictability by unveiling the black box and revealing what features have the most impact on classification, which is especially important in ensuring the stability and safety of real-time control systems.