A hybrid deep learning model for user story effort estimation

Area of Science:

• Software Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Accurate effort estimation in agile software development is challenging due to subjective requirements and natural language complexity.
• Existing methods often struggle with the nuances of user story descriptions.
• Data-driven approaches are needed to enhance estimation accuracy.

Purpose of the Study:

• To propose a hybrid Deep Learning (DL) model for data-driven effort estimation of user stories.
• To develop and validate a novel model using advanced semantic modeling on large-scale textual data.
• To improve the accuracy and reduce bias in agile software development effort estimation.

Main Methods:

• A hybrid DL model integrating Bidirectional Encoder Representations from Transformers (BERT) and Long Short-Term Memory (LSTM) was developed.
• A curated dataset of 4,079 user stories was created through systematic preprocessing and expert validation.
• The model was evaluated against ML baselines using a multi-metric framework and statistical tests (Wilcoxon Signed-Rank Test).

Main Results:

• The hybrid DL model demonstrated superior performance with MAE=0.6481 and RMSE=1.4559, significantly outperforming conventional methods.
• Mapping continuous outputs to the Fibonacci sequence achieved 72% classification accuracy for Scrum planning.
• Statistical validation confirmed significant performance improvements (p < 0.05).

Conclusions:

• The hybrid DL model effectively reduces estimation bias and enhances predictive quality in agile software development.
• A web-based decision support system was operationalized for real-time estimation.
• Future work includes integrating eXplainable AI (XAI) and validating on industrial datasets.