Software Fault Prediction through Hybrid Learning Algorithms with Integrated Change Metrics

Abstract

Software fault prediction (SFP) is essential for improving software reliability and reducing development costs by identifying fault-prone modules early. This paper investigates the integration of change metrics—such as code churn, commit frequency, and modification history—with hybrid learning algorithms to enhance prediction accuracy. We propose a novel hybrid framework that combines genetic algorithms (GA) for feature selection, convolutional neural networks (CNN) for extracting semantic and temporal features from change data, and multi-layer perceptrons (MLP) for processing traditional static metrics. These components are fused using a gated attention mechanism to optimize predictions. Empirical evaluations on datasets from the PROMISE repository, including projects like Ant, Camel, and Xerces, demonstrate that our approach achieves superior performance in terms of F1-score (0.87), AUC (0.90), and effort-aware metrics like PofB20 (0.55), outperforming existing hybrid models by 8-15%. The study addresses key challenges such as class imbalance, cross-project generalizability, and the underutilization of dynamic change data. By incorporating real-time evolutionary metrics, our model enables more proactive defect management, contributing to advancements in software engineering practices.