Deep Learning-Based Classification of Software Bugs Using Code Context and AST Features

Abstract

A software component plagued with bugs is likely to experience both functional and non-functional difficulties, such as usability, performance, and security issues. These bugs can range from improper layouts to system crashes and security vulnerabilities. In this research, we developed a deep learning-based model to classify software bugs based on preceding and corrected code statements. Preprocessing includes extracting features from the Abstract Syntax Tree (AST) by traversing the tree to capture node types and relationships, then encoding the AST into a numerical vector representation for classification. Using AST and padding methods, features are extracted from code statements for use in a Convolutional Neural Network (CNN) model. We proposed a CNN-based model with AST-based feature extraction for large-scale software bug classification and evaluated on 153,652 bug samples from 1,000 GitHub projects. Compared to traditional datasets, this large dataset presents challenges in building accurate prediction models. This model is particularly useful in continuous integration (CI) pipelines, where it can automatically detect problematic code during the build process, helping to identify bugs faster and reduce manual review.