A Deep Generative Ensemble Framework Based on Tabular Variational Autoencoder for Imbalanced Software Defect Prediction

Moch. lutfi; Muhammad Faishol Amrulloh; Zulfatun Nikmatus Saadah; Ahmad Zulham Fahamsyah Havy; Arief Tri Arsanto

doi:10.5935/jetia.v12i59.3339

Authors

Moch. lutfi Computer Science, Universitas Yudharta Pasuruan, Jl. Yudharta No.7, Kembangkuning, Sengonagun, Purwosari, Pasuruan, 67162, East Java, Indonesia. https://orcid.org/0000-0001-9227-759X (unauthenticated)
Muhammad Faishol Amrulloh Computer Science, Universitas Yudharta Pasuruan, Jl. Yudharta No.7, Kembangkuning, Sengonagun, Purwosari, Pasuruan, 67162, East Java, Indonesia. https://orcid.org/0009-0009-9721-1480 (unauthenticated)
Zulfatun Nikmatus Saadah Computer Science, Universitas Yudharta Pasuruan, Jl. Yudharta No.7, Kembangkuning, Sengonagun, Purwosari, Pasuruan, 67162, East Java, Indonesia https://orcid.org/0009-0004-2450-7268 (unauthenticated)
Ahmad Zulham Fahamsyah Havy Computer Science, Universitas Yudharta Pasuruan, Jl. Yudharta No.7, Kembangkuning, Sengonagun, Purwosari, Pasuruan, 67162, East Java, Indonesia. https://orcid.org/0009-0009-3006-7963 (unauthenticated)
Arief Tri Arsanto Computer Science, Universitas Yudharta Pasuruan, Jl. Yudharta No.7, Kembangkuning, Sengonagun, Purwosari, Pasuruan, 67162, East Java, Indonesia. https://orcid.org/0009-0001-2386-8940 (unauthenticated)

DOI:

https://doi.org/10.5935/jetia.v12i59.3339

Abstract

Context: Class imbalance is a prevalent issue in Software Defect Prediction (SDP), where non-defective modules significantly outnumber defective ones. This imbalance often degrades the performance of prediction models. While oversampling techniques like SMOTE are commonly used to address this problem, they often produce synthetic instances that are either too similar to the original data (leading to overgeneralization and high false positive rates) or overly diverse (reducing the model’s ability to detect defects).

Objective: This study proposes a novel hybrid approach that combines a generative oversampling method, Tabular Variational Autoencoder (TVAE), with ensemble learning to improve defect prediction performance. The aim is to achieve a better balance between precision and recall while effectively handling imbalanced datasets.

Method: TVAE is used to generate realistic synthetic defect instances by learning the underlying joint distribution of tabular software metrics. The synthetic data is then used to train ensemble models, including TVAE+Naïve Bayes with Bagging and TVAE+Random Forest with Gradient Boosting Stacking. The proposed methods are evaluated on seven NASA PROMISE datasets and compared against baseline and state-of-the-art approaches using AUC, G-Mean, Accuracy, and F1-Score.

Results: Experimental results show that TVAE+RF+GB consistently outperforms other methods, achieving an average accuracy of 93%, AUC of 0.91, and F1-Score of 0.93. TVAE+NB+Bagging also demonstrates significant improvement over baseline models, particularly in G-Mean. Statistical tests confirm the significance of these improvements in most evaluation metrics.

Conclusion: The integration of TVAE with ensemble learning effectively addresses class imbalance in SDP and enhances prediction performance. TVAE+RF+GB is recommended as a robust and efficient solution for imbalanced defect prediction datasets.