Transformer-Enhanced Deep Characterization of Sprints in Agile Software Development

Abstract: Agile methodologies, particularly Scrum, have become foundational in modern software development, with sprints serving as time-boxed iterations for incremental delivery. Accurate sprint outcome prediction is essential for effective project management and resource allocation. While Sprint2Vec introduced a deep learning framework using LSTM networks to characterize sprints through vector embeddings, it faces limitations in modeling long-range dependencies due to sequential processing constraints.

This paper presents Sprint2Vec +, which extends Sprint2Vec by integrating transformer-based architectures that leverage self-attention mechanisms to capture complex interactions among sprint elements. We evaluate our approach using datasets from five open-source projects—Apache, Atlassian, Jenkins, Spring, and Talendforge — containing over 5,000 sprints and 71,000 issues. Sprint2Vec + demonstrates a 10.8% improvement in combined prediction accuracy (10.3% for productivity, 11.3% for quality) with 40.9% faster inference time compared to the original LSTM-based approach. Beyond predictive improvements, Sprint2Vec + provides interpretable attention patterns revealing critical relationships between planning activities and implementation outcomes. Statistical significance testing (p < 0.001) confirms the robustness of improvements across all project domains. This work contributes a scalable, transformer-based solution for advancing sprint analytics in Agile software engineering.