Scholarlink

<p>Accurately predicting whether a software development task can be completed within a short timeframe, such as one week, is critical for effective agile project management. Traditional methods and standard machine learning models often struggle with the dynamic, heterogeneous nature of software tasks, particularly the evolving textual information and irregular update intervals. Many predictive models fail to create a holistic task representation, neglecting the complex and time-sensitive interactions between features such as text comments and task priorities, as well as ignoring the irregular, dynamically updated time. To address this, a novel predictive framework is introduced that learns a context-aware task representation by integrating various features: textual, categorical, and numeric. Unlike standard Transformer architectures that rely on fixed positional encodings, the proposed architecture employs a feature-specific time-aware attention mechanism to explicitly capture the varying impact of update recency in irregular sequences. This is combined with multi-channel embedding and a hierarchical multi-head attention mechanism to model complex interdependencies both within and across feature modalities. Experimental results on a dataset comprising 37 real-world software projects show that the proposed model attains higher performance in F1-score, precision, recall, accuracy, and AUC-ROC. The model outperforms several strong baseline methods, including ANN, XGBoost, LSTM, a standard Vanilla Transformer, and a fine-tuned RoBERTa model. The robustness of these results is further confirmed through a 5-fold cross-validation analysis, demonstrating the model’s stability across different project contexts. Furthermore, a detailed ablation study isolating specific architectural components confirms the necessity of the time-aware attention and hierarchical fusion mechanisms. A comparative efficiency analysis demonstrates that the proposed framework achieves a superior balance between predictive accuracy and computational feasibility for real-time applications compared to LLM-based approaches. A new qualitative case study demonstrates how the model identifies hidden “blocker” comments to trigger proactive management interventions, validating the practical utility of the system.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Learning context-aware representation for short-term software task completion prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Dong Wu,
  • Michael Grenn

摘要

Accurately predicting whether a software development task can be completed within a short timeframe, such as one week, is critical for effective agile project management. Traditional methods and standard machine learning models often struggle with the dynamic, heterogeneous nature of software tasks, particularly the evolving textual information and irregular update intervals. Many predictive models fail to create a holistic task representation, neglecting the complex and time-sensitive interactions between features such as text comments and task priorities, as well as ignoring the irregular, dynamically updated time. To address this, a novel predictive framework is introduced that learns a context-aware task representation by integrating various features: textual, categorical, and numeric. Unlike standard Transformer architectures that rely on fixed positional encodings, the proposed architecture employs a feature-specific time-aware attention mechanism to explicitly capture the varying impact of update recency in irregular sequences. This is combined with multi-channel embedding and a hierarchical multi-head attention mechanism to model complex interdependencies both within and across feature modalities. Experimental results on a dataset comprising 37 real-world software projects show that the proposed model attains higher performance in F1-score, precision, recall, accuracy, and AUC-ROC. The model outperforms several strong baseline methods, including ANN, XGBoost, LSTM, a standard Vanilla Transformer, and a fine-tuned RoBERTa model. The robustness of these results is further confirmed through a 5-fold cross-validation analysis, demonstrating the model’s stability across different project contexts. Furthermore, a detailed ablation study isolating specific architectural components confirms the necessity of the time-aware attention and hierarchical fusion mechanisms. A comparative efficiency analysis demonstrates that the proposed framework achieves a superior balance between predictive accuracy and computational feasibility for real-time applications compared to LLM-based approaches. A new qualitative case study demonstrates how the model identifies hidden “blocker” comments to trigger proactive management interventions, validating the practical utility of the system.