Envisioning the Future of Debugging: The Advent of ABERT for Adaptive Neural Localization of Software Anomalies

Locating bugs is a crucial but difficult aspect of program maintenance. An improved machine learning approach called Adaptive Bidirectional Encoder Representations from Transformers (ABERT) is presented in this work to improve the conventional bug localization procedure. ABERT utilizes an adaptive attention mechanism and dynamic depth adjustment influenced by bug reports and code complexity. Using various datasets, we compared common machine learning models such as Logistic Regression, Random Forest, Support Vector Machines, and Gradient Boosting. ABERT improved performance by achieving greater accuracy, precision, Recall, F1 score, and AUC metrics. Clustering analysis was conducted to uncover inherent data groups, improving our comprehension of bug report features. The study on cross-dataset generalization highlighted ABERT's strong and flexible characteristics. The results support sophisticated machine learning methods in bug localization, indicating a notable transition towards smarter and automated software development tools.