Abstract
In large-scale software engineering, managing dependencies between thousands of tasks, bugs, and user stories is a significant challenge. Manual identification of these links is labor-intensive and error-prone, leading to fragmented project data. This paper introduces TaDeR (Task Dependency Recommendation), a deep learning-based approach that utilizes Siamese Neural Networks and 1D Convolutional Neural Networks (CNNs) to automatically suggest task dependencies. By incorporating temporal features and a specialized time-based filter, TaDeR achieves superior performance compared to traditional keyword-matching techniques. Evaluation on real-world datasets from the Apache Flume and Moodle communities demonstrates the effectiveness of the proposed model in identifying relevant task links within project backlogs.
1. Introduction
Modern software development relies heavily on project management platforms like Jira to track progress. As projects scale, backlogs often grow to encompass thousands of individual items. Identifying dependencies—such as when one task blocks another or when two issues are related—is crucial for maintaining a coherent development roadmap. However, the "needle in a haystack" nature of searching through massive backlogs means many dependencies are never formally linked. This lack of visibility results in duplicated efforts and misaligned priorities.
TaDeR addresses this by treating dependency identification as a binary classification problem. Given a pair of tasks, the system calculates the probability that a link should exist between them, providing developers with automated recommendations to streamline project organization.
2. Methodology
The TaDeR framework is built upon a Siamese Neural Network architecture designed to process and compare pairs of task descriptions.
2.1 Siamese Neural Network Architecture
The core of the system consists of two identical subnetworks that share weights. Each subnetwork processes a task's textual data (Title, Description, and Summary) to generate a high-dimensional feature vector. The outputs of these subnetworks are then compared to determine the similarity between the two tasks.
The authors selected 1D Convolutional Neural Networks (CNNs) for feature extraction. While Recurrent Neural Networks (RNNs) like LSTMs are common in natural language processing, the researchers found that CNNs were more effective for technical task data. Technical reports often rely on specific keywords and discrete jargon rather than complex narrative structures. CNNs excel at capturing these local, position-invariant features, making them more robust for this specific domain.
2.2 Word Embeddings
To represent text numerically, TaDeR utilizes GloVe (Global Vectors for Word Representation). GloVe was chosen over alternatives like FastText because its focus on global co-occurrence statistics produced more discrete word vectors in the feature space. This characteristic proved beneficial for distinguishing between the highly specific technical terms found in software issue trackers.
2.3 Temporal Features and the Time Filter
A significant contribution of this work is the integration of temporal data. The model incorporates two specific time-based features:
- CC (Creation-Creation): The time interval between the creation dates of the two issues.
- CU (Creation-Update): The time interval between the creation of one issue and the last update of the other.
Furthermore, the system employs a Time Filter during post-processing. Based on the observation that related tasks are often created in temporal "bursts" (e.g., during a specific sprint or feature push), the filter limits candidate recommendations to tasks created within a specific window, such as one month. This drastically reduces the search space and filters out irrelevant noise.
3. Experimental Evaluation
The researchers evaluated TaDeR using two large-scale, real-world datasets:
- FLUME: 3,373 issues from the Apache Flume project.
- MDLSITE: 5,910 issues from the Moodle Community Sites.
To ensure the evaluation reflected real-world usage, the data was split chronologically. The model was trained on historical data and tested on its ability to predict links for subsequent, newer tasks.
4. Results and Discussion
TaDeR consistently outperformed traditional baseline methods, such as TF-IDF combined with Cosine Similarity.
4.1 Performance Metrics
On the FLUME dataset, the GloVe + CNN configuration achieved an Accuracy@1 of 0.2335. The model's utility is even more apparent in its Top-K performance. For the MDLSITE dataset, the system reached an Accuracy@5 of 0.5270, indicating that a correct dependency is found within the top five suggestions more than half the time.
4.2 Key Findings
- Feature Synergy: The best results were achieved by combining the Title, Description, and Summary (DTS). Using the Title alone yielded poor results, highlighting the necessity of the rich context provided in the full task description.
- Temporal Impact: The introduction of the Time Filter was a critical factor in boosting precision, confirming that temporal proximity is a strong indicator of task relevance in software development.
5. Conclusion and Future Work
TaDeR demonstrates that deep learning, specifically Siamese CNNs combined with temporal filtering, can significantly improve task dependency recommendation in project management. While the current system effectively identifies the existence of a link, future work will focus on Link Classification—identifying whether a link represents a duplicate, a sub-task, or a blocking dependency. Additionally, the authors plan to investigate Transformer-based models like BERT to further enhance the semantic understanding of technical requirements.
Summary updated