AI-Based Automated Software Testing Model for Predicting Bug-Prone Modules in Large Software Systems

Abstract

Predicting software bugs in software development life cycle is a crucial part of improving software quality. Conventional
methods of testing can be slow and cannot be used in large-scale systems, necessitating intelligent and automated methods.
In the current research, a hybrid model of LightGBM and GRU is suggested, which would contribute to the improvement of
defect prediction. The dataset of NASA PROMISE repository JM1 is utilized, and it includes software metrics of complexity
and structure of the code. Preprocessing of the dataset involves such steps as data cleaning, data outlier, feature scaling
and class imbalance correction using SMOTE. LightGBM model is also useful in capturing feature interactions in structured
data but the GRU model learns complex non-linear patterns. The two models are used to produce final predictions by
combining the output of both models through a soft voting mechanism. Through experiments, the hybrid model is found
to have high accuracy of 95.11 percent and it is superior to current models (LSTM, Random Forest). The results reveal
that the suggested solution is a powerful and effective solution in predicting bug-prone modules, which finally assists to
increase the reliability of the software and optimization of the testing process.