Learning Context-Aware Representation for Short-Term Software Task Completion Prediction
Open Access Depositedfor operational use, the same output can also be interpreted as a complementary delay-risk score. Overall, the results support the framework as a reliable research prototype for agile decision support, while also indicating that deployment in a new organizational environment would require project-specific validation, threshold calibration, and domain adaptation.
Accurately predicting whether a software development task is likely to be completed withina short horizon, here set to 7 days, is critical to agile project management. Classical estimation approaches and traditional machine-learning models often have difficulty representing the heterogeneous and unevenly timed evidence found in software tasks, including evolving text, structured metadata, and irregular update intervals. To address this problem, this dissertation evaluates a context-aware predictive model that combines textual, categorical, and numerical variables. While typical Transformers rely on static positional encoding, the proposed model uses feature-specific time-aware attention layers to capture the recency of irregular task updates. This mechanism is combined with multi-channel embeddings and hierarchical multi-head attention modules to model dependencies within and across feature modalities. Under the reported evaluation settings on 37 Apache Jira projects, and relative to the evaluated baselines, the proposed model achieves the strongest average performance in F1-score, precision, recall, accuracy, ROC-AUC, and PR-AUC for the 7-day prediction task. The results are supported by held-out project evaluation, projectlevel cross-validation, training-partition statistical comparison, chronological evaluation, ablation analysis, duration-stratified analysis of short-cycle and long-running tasks, and secondary validation on GitHub issues. The duration analysis keeps the 7-day prediction target fixed and shows that the proposed architecture reaches its strongest duration-specific performance on tasks completed between 3 and 7 days. Tasks completed within 3 days are harder because the issue tracker often contains too little evolving evidence before closure, while tasks extending beyond 21 days remain challenging because their histories are longer and noisier. Taken together, these analyses show that the time-aware attention mechanism and hierarchical multichannel fusion design make a substantive contribution to predictive performance under the reported settings, while also clarifying that the model is best aligned with task lifecycles near the 7-day decision window. The GitHub validation shows that the model remains competitive beyond the Apache Jira setting, while also indicating that domain adaptation is needed before use in a substantially different ecosystem. Comparative efficiency analysis further distinguishes the full SBERT-including pipeline from the taskspecific classifier with precomputed embeddings, showing that the latter is smaller and faster than the fine-tuned RoBERTa baseline under the reported setup. The model predicts whether a task will be completed within 7 days
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.