CorrBoost: a feature selection technique and utility of tabular deep neural networks in software fault prediction

Abstract

Context: Software Fault Prediction (SFP) leverages supervised Machine Learning to detect faulty software constructs using software metrics and corresponding labels. Despite recent advances in Deep Learning (DL) for tabular data, their application to SFP remains underexplored. Objectives: This study proposes a novel feature selection method, CorrBoost, which combines correlation analysis and XGBoost to address feature dimensionality. Additionally, we evaluate existing tabular DL architectures, super convergent deep neural networks (sDNN) and TabNet for SFP. Methods: Using 26 public datasets from NASA, PROMISE, and AEEEM repositories, we apply the adaptive synthetic oversampling technique to manage class imbalance. We compare DL models with five state-of-the-art techniques and two gradient-boosted tree models (XGBoost and LightGBM) using AUC-ROC, AUPRC, and Accuracy. Statistical significance is validated using the Bayesian Signed Rank Test and Scott-Knott ESD. Results: Gradient-boosted trees and existing state-of-the-art models outperform DL methods in AUC-ROC by 17.9% and 9.6%, respectively. CorrBoost achieves a 55% average reduction in feature dimensionality with negligible performance loss. DL methods, however, incur significantly higher processing time and perform poorly on unseen test data. Conclusion: CorrBoost combined with boosted tree models offers a superior trade-off between performance and computation. While tabular DL architectures hold promise, they currently lag behind traditional methods for SFP on real-world data.