Stacked BiLSTM-GRU Deep Learning Model Improves Cross-Project Software Fault Prediction

A stacked deep learning framework combining Bidirectional Long Short-Term Memory units with Gated Recurrent Units has been developed to predict software faults across projects by strengthening the modeling of temporal dependencies in software metrics. Designed to identify error-prone modules early in the software development process, the architecture aims to enhance system reliability and control ongoing maintenance demands.

The hybrid framework was evaluated using six established PROMISE benchmark datasets, consisting of JEdit, Ant, Camel, Log4J, Xalan, and Lucene. Across all evaluated datasets, the combined architecture consistently outperformed conventional machine-learning techniques as well as standalone recurrent models.

Across the test datasets, the model recorded an average Mean Squared Error of 0.0181, a Mean Absolute Error of 0.0614, an Average Absolute Error of 0.0579, and an Average Relative Error of 0.0321. The evaluation also demonstrated an average pred(l) accuracy of 98.81 percent, reaching 99.25 percent on the Log4J repository, alongside an average coefficient of determination of 0.961.

Statistical significance across all comparative assessments was verified using the Wilcoxon signed-rank test, yielding p-values below 0.0004. Cross-project experiments also demonstrated competitive performance, indicating that the stacked design offers a reliable and practical approach for fault prediction across diverse software engineering contexts.

Subject of Research: Cross-project software fault prediction using stacked recurrent deep learning models

Article Title: Cross-project software fault prediction using a stacked BiLSTM-GRU model: statistical and empirical validation

Article References: Jha, S. K., Hati, A. N., Rathore, N. C., Panda, A. K., Kumar, S., & Anand, K. (2026). Cross-project software fault prediction using a stacked BiLSTM-GRU model: statistical and empirical validation. Cluster Computing, 29(16), Article 933. https://doi.org/10.1007/s10586-026-06653-1

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06653-1

Keywords: software fault prediction, deep learning, BiLSTM, Gated Recurrent Units, PROMISE datasets, cross-project prediction, system reliability, Cross-project, software, fault, prediction, stacked