An explainable parallel multi-transformer ensemble framework for agile software effort estimation
Effort estimation of user stories remains a challenge in the agile software development. Inaccurate estimation can cause schedule delays and cost overruns. This research proposed a novel explainable Parallel Multi-Transformer Ensemble (PMTE) Framework to improve the accuracy of user story effort estimation. The main contribution of this work is the curation of a dataset of 4079 unique user stories that were gathered from various open source software projects. To overcome the problem of data noise and leakage industrial scrum experts designed and developed a three-stage curation protocol to curate and validate the dataset. The Design Science Research (DSR) methodology was followed to design a hybrid parallel ensemble framework that combines two pre-trained transformer encoders, Bidirectional Encoder Representations from Transformers (BERT) and Distilled Robustly Optimized BERT Approach (DistilRoBERTa), with an optimized Term Frequency-Inverse Document Frequency (TF-IDF) keyword frequency branch. These fused semantic-frequency features are then fed into a bidirectional Long Short-Term Memory (BiLSTM) network and projected onto the regularized regression head. Various standard regression metrics were used to assess framework performance. Experimental results demonstrate that the proposed framework is able to achieve high predictive performance, obtaining a continuous Mean Absolute Error (MAE) of 0.5124, an agile Fibonacci-mapped MAE of 0.4443 and R 2 of 0.9283 on the independent test set, which shows superior performance in learning directly from unstructured text. We tested the framework on an external, independent set of 514 user stories from a private software house, and the results were stable, thus ensuring the generalizability of the framework. To address the black-box issue of deep neural networks, post-hoc eXplainable AI (XAI) was integrated at the word-level, providing project managers with clear and visual understanding of the model’s predictions. This framework translates these insights into a web-based decision support tool that is reliable, objective, and interpretable for Scrum sprint planning and backlog grooming. The results show that using parallel semantic representations and local keyword frequencies can significantly improve the accuracy of the estimation, providing a reliable, objective, and transparent data-driven baseline that helps the Scrum team plan its activities.
Authors
- Saadia Malik (ORCID: https://orcid.org/0000-0002-7057-4160)
- Muhammad Hamid (ORCID: https://orcid.org/0000-0002-2440-6596)
- Muhammad Saleem
Institutions
- King Abdulaziz University (SA)
Publication Details
- Journal
- Journal Of Big Data
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1186/s40537-026-01602-6
- Primary Topic
- Software Engineering Techniques and Practices
- Type
- article
- Field-Weighted Citation Impact
- 0.00