Automated Software Engineering最新文献

Accurate bug classification is essential for improving software quality, particularly in the emerging and complex domain of quantum computing. This paper introduces a rule-based framework for automated classification of quantum software issues across five dimensions: bug type, bug category, severity, quality attribute, and quantum-specific subtype. The framework integrates weighted keyword heuristics, TF–IDF similarity, and domain-specific rules to capture both general software defects and quantum-domain failure modes. The proposed approach was applied to 12,910 issues from 36 Qiskit repositories and validated on a stratified subset of 4,984 manually annotated issues. On this manually labeled subset, the framework achieved accuracies between 0.82 and 0.85 and macro-F1 scores ranging from 0.68 to 0.77, demonstrating strong agreement with human annotations without requiring supervised training. When compared with standard machine-learning baselines (Logistic Regression, Decision Tree, Random Forest, Gradient Boosting), the rule-based approach consistently outperformed all models across tasks, showing particularly large improvements in fine-grained and low-frequency categories such as Category (macro-F1: 0.26 vs. 0.69) and Quantum-Specific Subtype (0.15 vs. 0.77). Beyond predictive accuracy, the framework was applied to real-world 12,910 issues from 36 Qiskit repositories for large-scale distributional analysis. The results revealed that approximately 67% of issues were classical and 27% quantum-specific, with interoperability, usability, and maintainability identified as the most frequently affected quality attributes. Low-severity issues dominated (68.8%), while critical bugs accounted for around 21%. Quantum-specific defects were most prevalent at the circuit and gate abstraction levels, reflecting the hybrid and hardware-constrained nature of current quantum software development. Overall, the proposed rule-based framework offers a transparent, interpretable, and empirically validated approach for automated bug triaging in quantum software. Beyond its immediate practical utility, it provides a reproducible methodological framework that can support future hybrid and learning-based advances in quantum software engineering.

The success of technical Q&A sites such as Stack Overflow depends on two key factors: (a) active user participation and (b) the quality of the shared knowledge. Stack Overflow introduced an edit system that allows users to suggest improvements to posts (i.e., questions and answers) to enhance the quality of the content. However, users, such as post owners or site moderators, can reject these suggested edits by rollbacks due to unsatisfactory or low-quality edits, or because they violate edit guidelines. Unfortunately, subjectivity bias in determining whether an edit is satisfactory or unsatisfactory can lead to inconsistencies in the rollback decisions. For example, one user might accept the formatting of a method name (e.g., getActivity()) as a code term, while another might reject it. Such inconsistencies can demotivate and frustrate users whose edits are rejected. Furthermore, several post owners prefer to keep their content unchanged and even resist necessary edits. As a result, they sometimes roll back necessary edits and revert posts to a flawed version, which violates editing guidelines. The problems mentioned above are further compounded by the lack of specific guidelines and tools to assist users in ensuring consistency in rollback actions. In this study, we investigate the types, prevalence, and impact of rollback edit inconsistencies and propose a solution to address them. The outcomes of this research are fivefold. First, we manually investigated 764 rollback edits (382 questions + 382 answers) and identified eight types of inconsistent rollback. Second, we surveyed 44 practitioners to assess the impact of rollback inconsistencies. More than 80% of the participants found our identified inconsistency types detrimental to post quality. Third, we developed rule-based algorithms and machine learning (ML) models to detect the eight types of rollback inconsistencies. Both approaches achieve over 90% accuracy. Fourth, we introduced a tool, iEdit, which integrates these algorithms into a browser extension and assists Stack Overflow users during their edits. Fifth, we surveyed 16 Stack Overflow users to evaluate the effectiveness of iEdit. The participants found the tool’s suggestions helpful in avoiding inconsistent rollback edits.

In deep neural network (DNN) testing, high-quality test cases are essential to ensure the effectiveness and completeness of the testing. Mutation testing has been used in DNN testing to evaluate the quality of test cases. However, existing methods don’t consider the importance of different neurons; they randomly select neurons for mutation, resulting in low-quality mutants. To address this issue, we propose a novel method for generating high-quality mutants tailored for DNNs. Our approach first analyzes the output distribution of neurons and calculates their activation thresholds. Depending on these thresholds, we identify the behavior of neurons and select the most frequently activated neurons for each category as category-critical neurons. Next, we mutate these category-critical neurons to produce a candidate set of high-quality mutants. Although mutating category-critical neurons could reveal more errors, some mutants can cause significant changes in DNNs, making them easily killed and reducing the effectiveness of mutation testing. Thus, we further optimize the candidate set to get high-quality mutants. We calculate each mutant’s Category Above Ratio (CAR) score according to its execution behavior. Lastly, mutants whose CAR score is greater than or equal to the CAR threshold will be selected as higher-quality mutants. We also validate our method on several popular datasets and models. Experimental results demonstrate that our approach generates higher-quality mutants than random methods, reducing the number of mutants by 79.7% to 96.3%, while maintaining the effectiveness of mutation testing. Furthermore, retraining models with test cases generated by high-quality mutants could improve the robustness of the models.

Textual Vulnerability Descriptions (TVDs) are crucial for security analysts to understand and address software vulnerabilities. However, the key aspect inconsistencies in TVDs from different repositories pose challenges for achieving a comprehensive understanding of vulnerabilities. Existing approaches aim to mitigate inconsistencies by aligning TVDs with external knowledge bases, but they often discard valuable information and fail to synthesize comprehensive representations. In this paper, we propose a domain-constrained LLM-based synthesis framework for unifying key aspects of TVDs. Our framework consists of three stages: 1) Extraction, guided by rule-based templates to ensure all critical details are captured; 2) Self-evaluation, using domain-specific anchor words to assess semantic variability across sources; and 3) Fusion, leveraging information entropy to reconcile inconsistencies and prioritize relevant details. This framework improves synthesis performance, increasing the F1 score for key aspect augmentation from 0.82 to 0.87, while enhancing comprehension and efficiency by over 30%. We further develop Digest Labels, a practical tool for visualizing TVDs, which human evaluations show significantly boosts usability.

With the continuous development of digital learning environments, the demand for providing transparency and explanations about learner profiles (LPs) in the digital era is constantly increasing. Digital learning traces play a fundamental role in understanding LP categories and justifying the reasoning behind their behavior, such as learners in difficulty, active/inactive learners, those in progress, and success-oriented vs. at-risk learners. However, effectively identifying LPs from the massive data generated by these environments is challenging due to its vulnerability to misunderstandings and biases. Moreover, stakeholder may not know what queries to ask or what patterns to look for in the data, while traditional Learning Analytics tools struggle to bridge the gap between granular digital traces and human-interpretable conceptual insights. To address this gap, we propose XAI-Profile, a framework that leverages Explainable AI (XAI) to transform raw learner data into interpretable, actionable profiles. XAI-Profile combines explainable machine learning models with interactive visual workflows to bridge between low-level data traces and high-level pedagogical insights. The framework is built on three pillars: (1) Goal-Oriented Requirements for capturing stakeholders’ needs by refining a problem into sub-problems using goal-oriented AND/OR refinement, and mapping them to the target LP with their data indicators. (2) Visual Analytics Design for mapping LP categories by automatically correlating them with trust objects (e.g., rule-based models, frequent sequences in the data), visualization goals, and analysis types. (3) Interaction and Actionability, a guided process enabling stakeholders to discover interesting and unexpected patterns, validate hypotheses, and drill into context-specific explanations. A case study using data from the écri+ project demonstrates how stakeholders can quickly identify high-level behavioral patterns and drill down into specific profiles of interest. The results highlight how XAI-Profile advances learning analytics practices by making AI-driven insights both trustworthy and actionable fostering a human-centered approach to learner analytics.

Software refactoring is essential for maintaining the quality and extendibility of source code. Although prior research achieved high accuracy in predicting software refactoring opportunities using traditional machine learning methods, further refinements can still boost predictive reliability across multiple refactoring operations. The main objective of this study is to develop a more accurate model for software refactoring predictions. The main objective of this study is to develop a more accurate model for software refactoring predictions. We propose a two-level Tree-based Stacking (TB-Stacking) Ensemble designed to predict refactoring opportunities at the class, method, and variable levels. Various stacking ensembles were built to identify the optimal configurations of base and meta-models. Additionally, two feature selection techniques were employed to identify the most relevant features for accurate predictions. We evaluated our model using a publicly available dataset comprising over two million refactoring instances. The performance of nine traditional machine learning models and seven tree-based ensembles was also assessed, with statistical comparisons drawn against our proposed TB-Stacking model. Results indicate that the TB-Stacking Ensemble consistently outperformed traditional models in all refactoring operations, and demonstrated competitive or superior performance compared to tree-based ensembles. It demonstrated significant improvements in prediction capabilities highlighted by statistical analysis. The model’s robust performance across various refactoring tasks establishes a new benchmark for future refactoring tools.

The growing complexity of software systems and the increasing demand for global software development necessitate innovative approaches to enhance project management, quality assurance, and risk mitigation. In this study, Artificial Intelligence (AI) and Machine Learning (ML) techniques are examined for overcoming problems in software project management, notably effort estimation, scheduling, resource allocation, risk management, and defect prediction. By systematically reviewing the literature, we show that AI/ML models like Support Vector Machines, neural networks, and ensemble learning can enhance estimation accuracy, maximize resource utilization, and reduce risks. Furthermore, the practical benefits and challenges of implementing an AI/ML system into a real-world system are discussed using real-world case studies, which include data quality and integration issues, and the interpretability of the model. In addition, advanced models, such as graph convolutional networks and deep neural networks, hold great promise as a defect prediction and bug severity classifier. The focus of this research is to leverage the transformative capabilities of AI/ML toward defect-free, efficient, and customer-centric software development. Finally, it suggests future research interests, including integrating explanation model AI, managing data in a better way, and implementing the scalable hybrid approach to meet the newer needs of the industry.

Adware represents a pervasive threat in the mobile ecosystem, yet its inherent characteristics have been largely overlooked by previous research. This work takes a crucial step towards demystifying Android adware. We present AdwareZoo, a comprehensive dataset comprising 15,996 adware samples across 118 distinct families collected from security reports and app repositories. We identify adware family payloads by isolating packages from samples for VirusTotal rescanning, unveiling distinctive patterns in family naming conventions, and exposing the misclassification of legitimate ad networks as adware. Our analysis of payload location strategies reveals that over 30% of adware families employ payloads beyond conventional Java/Kotlin code. Based on our dataset analysis, we conducted a comprehensive Adware Characterization of 92 distinct adware families, revealing diverse implementation patterns and evolving techniques across the mobile ecosystem. To facilitate this analysis, we developed an Adware Characterization Schema that provided a structured taxonomy for systematically classifying the observed behaviors. Our investigation uncovered multiple categories of fraudulent activities, including aggressive ad display techniques, sophisticated click fraud implementations, privacy information leakage, malicious promotion mechanisms, and various persistence and evasion mechanisms employed to avoid detection while maximizing illicit revenue. This research establishes foundations for comprehending the fraudulent and adversarial techniques within the mobile adware landscape and facilitates the development of more robust detection mechanisms against these evolving threats.

Most studies focused on information retrieval-based techniques for fault localization, which built representations for bug reports and source code files and matched their semantic vectors through similarity measurement. However, such approaches often ignore some useful information that might help improve localization performance, such as 1) the interaction relationship between bug reports and source code files; 2) the similarity relationship between bug reports; and 3) the co-citation relationship between source code files. In this paper, we propose a novel approach named Multi-View Adaptive Contrastive Learning for Information Retrieval Fault Localization (MACL-IRFL) to learn the above-mentioned relationships for software fault localization. Specifically, we first generate data augmentations from report-code interaction view, report-report similarity view and code-code co-citation view separately, and adopt graph neural network to aggregate the information of bug reports or source code files from the three views in the embedding process. Moreover, we perform contrastive learning across these views. Our design of contrastive learning task will force the bug report representations to encode information shared by report-report and report-code views, and the source code file representations shared by code-code and report-code views, thereby alleviating the noise from auxiliary information. Finally, to evaluate the performance of our approach, we conduct extensive experiments on five open-source Java projects. The results show that our model can improve over the best baseline up to 28.93%, 25.57% and 20.35% on Accuracy@1, MAP and MRR, respectively.