Cross‐project defect prediction (CPDP) is a promising solution for new software projects lacking historical data. However, modern privacy/compliance constraints and limited sharing budgets often restrict how many labeled source instances can be transferred across organizations. Existing instance selection strategies typically prioritize “safe” majority instances or rely on random sampling, which can miss boundary‐defining defects under very small budgets. To address this, we propose weakness‐aware subset selection (WASS), a framework for source‐side local curation under a source‐sharing budget. WASS combines target‐domain similarity, weakness‐aware scoring, and class‐aware quota allocation to curate a compact transferable subset of source instances. Extensive experiments on 88 directed same‐family CPDP pairs with seven diverse classifiers (including transfer component analysis [TCA] + logistic regression [LR] and support vector machine [SVM]) show a clear 1%–5% Golden Window under the evaluated protocol. The experimental protocol also clarifies source‐sharing accounting, repeated‐run randomness, and family‐level variation. Therefore, WASS is a practical, plug‐and‐play cold‐start accelerator for source‐sharing‐constrained CPDP.
Driven by the rapid advancement of un-manned aerial vehicle (UAV) and deep learning technologies, UAV surveillance has emerged as a vital supplement to conventional ground monitoring systems due to its flexible deployment, wide coverage, and real-time capability. However, most state-of-the-art human action recognition models are trained on ground-level video data and struggle to adapt to the inherent characteristics of aerial imagery, including extremely small object scales, low resolution, complex background interference, and severe mutual occlusion. To address these technical bottlenecks in UAV-based human detection, this paper presents a lightweight yet high-precision detection method, dubbed RDDR-YOLO, based on YOLOv8s. Extensive experiments are conducted on two public benchmark datasets, VisDrone2019 and TinyPerson. Results show that RDDR-YOLO achieves mAP50 values of 43.9% and 36.4% on the two datasets, outperforming the YOLOv8s baseline by 7.0% and 9.6%, respectively. Its human localization precision reaches 41.2%, yielding an improvement of 14.5%. Moreover, the parameter count is reduced by 67.6%, and the inference speed reaches 70 FPS. The proposed method substantially improves the detection accuracy of small-scale human targets in UAV scenarios while maintaining competitive real-time performance. It can serve as the 2D localization module in the first stage of a two-stage human action recognition framework, effectively supporting human pose estimation and motion capture in UAV-based scenarios. The method holds significant practical potential in smart city security, emergency search and rescue, crowd management, and other related fields.
Decision-making in software development involves balancing multiple technical and organizational factors. While multicriteria decision-making (MCDM) methods offer systematic approaches to such problems, they typically do not account for practitioners' emotional states, which may influence judgments. This study examines how affective states impact decision-making and proposes a framework integrating psychometric evaluation with MCDM techniques. We present a framework combining MCDM, specifically the best-worst method (BWM) for problem formulation, with an affective scale, specifically the International Positive and Negative Affect Schedule - Short Form (I-PANAS-SF), to estimate the individual's affective state and mitigate the potential negative influence it might have on their decision. An experiment was conducted to observe how the valence of the affective state (whether positive or negative) may impact the importance assigned to decision criteria. Evaluations were filtered to include inputs from participants exhibiting positive affect, and the resulting prioritization outcomes were compared with an external benchmark dataset. The results indicate that the framework produced different aggregated decisions depending on affective valence, with positive-affect filtering yielding distinct prioritization patterns. Whether these differences improve decision quality requires direct comparison with affect-agnostic baselines. The exploratory comparison with the benchmark dataset revealed variations in task valuation that may be associated with affective states. These findings demonstrate the feasibility of incorporating psychometric assessment into MCDM-based software engineering (SE) decisions while highlighting the need for further validation using larger and more diverse samples, as well as careful consideration of ethical and practical deployment challenges.
Agile philosophy emphasises flexibility, adaptability and human-centred software development. However, in medical software development, integrating regulatory compliance and safety-critical requirements within Agile methods introduces significant tension. Compliance and safety-critical are often treated as afterthoughts rather than integrated in a dynamic, Agile workflow, leading to costly rework and system failures. To address these challenges, this study employed a qualitative, multi-method research design spanning four phases. Two phases involved in-depth interviews with 30 Agile practitioners from the UK, Nigeria and India, recruited using snowball sampling. In the third phase, a novel practice-based process model was developed. To establish the model’s practice relevance and transferability, the model was evaluated with eight Agile practitioners in the UK and Nigeria in the fourth phase. To improve credibility, the model was further validated through a case study involving Nigerian practitioners managing a Canadian medical software project. Data from each phase were inductively analysed using techniques informed by grounded theory: open coding, memoing, constant comparison and theoretical saturation. The study identified 39 Agile practices used by practitioners for medical software development. The findings reveal three novel artefacts, the compliance backlog, safety clinical backlog and misuse backlog, and two novel ceremonies: internal clinical demoing and the compliance review. In addition, three specialised roles emerged: the medical scrum master, clinical tester and regulatory expert. These practices were not theorised a priori but emerged inductively from practitioners’ interviews. Practitioners reported that regulatory compliance and safety outcomes improve when regulatory and safety-critical requirements are embedded dynamically within Agile workflows rather than treated as external constraints. This research contributes a novel, empirically grounded taxonomy of Agile roles, ceremonies and artefacts that goes beyond merely proposing another process design; it presents an operationalised Agile workflow shaped by the real experiences of Agile practitioners in regulated medical software environments. Reflecting how compliance and safety activities are coordinated in practice, the study demonstrates the embedding of ’compliance and safety-by-design’ within Agile software development. We recommend using the model, as it includes novel practices for reconciling contextual influences, including weak regulatory implementation and siloed development practices.
Automated radiology report generation (RRG) aims to reduce radiologist workload while improving reporting consistency and clinical reliability. Despite its importance for diagnosis, manual reporting remains time-consuming and error-prone. Existing encoder-decoder models often struggle to align visual and textual representations, resulting in incomplete or inconsistent reports. Large language models (LLMs) provide strong linguistic fluency for automated RRG, but most existing approaches still follow a linear generation paradigm and lack explicit reasoning, self-evaluation, and structured clinical grounding. To address these limitations, we propose SMK-RRG, a self-critical multisource Knowledge framework for RRG. The proposed framework integrates structured clinical knowledge, pathology-aware visual evidence, and iterative self-refinement within a unified LLM-driven architecture. It is implemented through self-critical thought (SCT) prompting, which performs multistage candidate report generation, LLM-based self-critique, and CheXbert-guided clinical selection among candidate reports. By enabling critique-guided prompt evolution across refinement stages, SCT moves beyond conventional chain-based prompting and supports nonlinear, clinically grounded reasoning. Experiments on public chest X-ray benchmarks demonstrate that SMK-RRG produces more accurate, interpretable, and clinically consistent reports, highlighting the effectiveness of self-critique and structured information integration for reliable clinical reports.
Long-term action anticipation (LTA) aims to predict an ordered sequence of future verb-noun actions from a partially observed video. While this task serves as the foundation for embodied intelligence, anticipating physically feasible long-term actions remains a critical challenge. Existing methods, which operate in an open-loop manner, often hallucinate non-existent objects, violate object affordances, or disregard object states, as they lack explicit mechanisms to verify action feasibility against the physical environment. To address this, we propose FactCheck, a novel multi-agent collaboration framework that improves feasibility through a closed-loop “Observe-Plan-Verify” mechanism. FactCheck decomposes the complex LTA task into specialized roles: an Observer that recognizes historical actions from video observations and constructs a dual-form structured memory, comprising a History Action Abstract that captures high-level human intentions and environmental status, and a History Action Graph that encodes object states and temporal dependencies; a Planner that generates draft future actions conditioned on both low-level historical actions and high-level History Action Abstract; and a Verifier that rigorously validates the draft against the History Action Graph and refines infeasible actions. Extensive experiments on the EPIC-Kitchens-55 and EGTEA Gaze+ benchmarks demonstrate that FactCheck consistently outperforms state-of-the-art (SoTA) methods. Our work establishes a new paradigm for feasibility-aware LTA, effectively closing the loop of action recognition, action prediction, and action verification.
Software engineering promotes the reuse of software components to accelerate the development process; however, reusing third-party components in AI4SE’s software supply chain (SSC) can introduce vulnerability risks. Once a new vulnerability in a third-party component emerges, developers need to determine whether the current project is affected by the vulnerability, that is, assess the vulnerability reachability issue in the SSC, which requires a significant amount of manpower and resources for assessment. To address this issue, we propose VulFinder, a multi-agent-driven framework for automated vulnerability reachability validation. VulFinder begins by using static code analysis tools to construct function call paths between downstream applications and dependency vulnerability APIs. Leveraging a multi-agent mechanism comprising a distillator, discriminator, generator, and validator, VulFinder iteratively generates exploit tests for methods along the call graph, effectively validating vulnerability reachability by executing these tests on downstream applications. We evaluate the performance of VulFinder across different programming language ecosystems, which were constructed from vulnerability datasets for Java and Python. Experiments show that VulFinder achieves 21% accuracy improvement over the state-of-the-art tool VESTA and a 7% accuracy improvement over the popular baseline tool TRANSFER on the Java dataset and also demonstrates robust generalizability on the Python dataset, significantly reducing false positives and false negatives and delivering an average efficiency improvement of over 1.5 times. We recommend that researchers and practitioners use VulFinder in practice, which integrates the code comprehension capabilities of large language models (LLMs) with the multi-agent framework to reduce false alarms and missed alarms in the reachability of vulnerabilities in SSC.
Automated graphical user interface (GUI) testing is essential for ensuring mobile app quality. However, existing related methods lack deep GUI understanding and cannot segment specific functions for targeted testing, resulting in wasted test events and low efficiency. This study aims to leverage large language models (LLMs) to enhance Monkey-based GUI testing by enabling function-level segmentation and tailored parameter generation driven by user requirements. We propose GPT-Monkey, which integrates Monkey’s randomness with LLM’s understanding capability. It establishes global interface associations and cross-modal alignment for LLM-driven function segmentation, employs parameter-retrieval augmented generation (RAG) to guide tailored parameter generation, and adopts a dual-feedback mechanism to iteratively optimize testing. Experiments show that GPT-Monkey improves crash detection by 16.7% and efficiency by 34.5% over the optimal baseline, and achieves 95% function segmentation accuracy and uncovers 397 real-world crashes on 1000 Google Play apps. The results demonstrate that LLM-driven function segmentation and tailored parameter generation significantly enhance the precision and efficiency of automated GUI testing.
The U-Net model has demonstrated strong performance in the field of medical image segmentation. Moreover, several enhanced and improved versions of this model have emerged by incorporating transformer or MLP modules. However, these network models still face challenges in overcoming the limitations of linear modeling and the lack of interpretability. Based on the excellent performance of Kolmogorov-Arnold network (KAN) in terms of accuracy and interpretability, we propose a new architecture called KAN-U-Net. This architecture fuses the KAN network module into the U-Net model, allowing KAN-U-Net to inherit the original performance of U-Net while also fusing the nonlinear representation ability and interpretability of KAN networks. The experiments on multiple datasets demonstrate that the KAN-U-Net model outperforms in terms of precision and accuracy.
Modern data and machine learning (ML) platforms have become critical infrastructure for enterprise decision-making, yet many organizations continue to struggle with unreliable pipelines, poor data quality, fragmented feature engineering, and models whose behavior is difficult to explain or govern. While substantial progress has been made in data platform and MLOps tooling, most systems remain organized around technical layers rather than business domains, leading to unclear ownership, semantic drift, and fragile governance processes. This paper presents a domain-driven architecture for data and AI platforms that treats data products, pipelines, features, and models as first-class domain artifacts. The approach applies domain-driven analysis to the entire analytics and ML lifecycle, enabling governance, quality enforcement, and lineage by construction rather than by after-the-fact controls. We introduce the core system model, describe the key architectural components, and present a reference architecture for large-scale enterprise environments. The approach is illustrated using an end-to-end enterprise analytics and AI platform scenario. The result is a platform that is not only scalable but also reliable, explainable, and governable.
Online video anomaly detection (VAD) necessitates real-time inference without access to future temporal context. Existing training-free methods, while promising, often suffer from correlation bias, misclassifying infrequent but benign actions (e.g., running for exercise) as anomalies due to a lack of causal reasoning. In this paper, we introduce multi-agent counterfactual dialectics (MACD), a novel framework that reframes VAD from statistical outlier detection to causal falsification. MACD operationalizes this via a structured adversarial debate among three specialized agents: a Prosecutor for high-recall risk screening, a Skeptic that performs instruction-conditioned causal interventions to generate benign counterfactual explanations, and an Arbitrator that adjudicates the final verdict based on the “evidence gap.” To ground this reasoning in visual reality, we construct a visual-semantic world model using dynamic scene graphs (DSG) and set-of-mark (SoM) prompting, reinforced by a visual re-verification loop (VRL) and a contextual memory bank to mitigate hallucinations. Our “screen-debate-decide” protocol ensures computational efficiency by reserving expensive dialectics only for ambiguous frames. Extensive experiments on UCF-Crime and XD-Violence demonstrate that MACD achieves state-of-the-art zero-shot performance.
Managing large-scale data, particularly those with high-dimensional features, poses a significant challenge in data analysis. With the rapid growth in data generation and the increasing importance of data analysis for researchers, extracting meaningful patterns from big data has become a critical concern. Data analysts utilize data mining techniques to discover knowledge, extract patterns, and generate rules. However, the high number of features often leads to substantial computational costs, requiring advanced hardware resources. Moreover, the ability to derive interpretable classification rules in high-dimensional big data systems is essential. This paper proposes a method for knowledge extraction from big data using rule-based granular computing (GrC). Initially, the architecture extracts preliminary fuzzy rules using GrC techniques and fuzzy clustering. Subsequently, it employs ensemble learning to generate optimized Takagi-Sugeno-Kang (TSK) rules. The results indicate that effective feature selection through fuzzy clustering significantly contributes to the generation of optimal rules. Furthermore, experimental outcomes demonstrate that the proposed model achieves notable performance improvements compared to conventional algorithms such as Random Forest and standard gradient Boosting.
Human action recognition (HAR) is a fundamental component of ubiquitous computing, yet its wide-range applications are hindered by privacy concerns. Specifically, high-accuracy models typically require cloud-based processing that compromises sensitive visual data, while privacy-preserving on-device models suffer from limited reasoning capacities and frequent hallucinations. To resolve this conflict, we introduce multiagent debate for HAR (MAD-HAR), a novel framework designed for strictly local environments. MAD-HAR leverages a lightweight vision-language model (VLM) with a granular prompt to convert visual inputs into semantic captions, anonymizing data before inference. To mitigate reasoning failures, a heterogeneous ensemble of N=7 diverse small and medium language model agents (ranging from 8B to 14B parameters) engages in a structured multiround debate. Rather than outputting simple labels, agents are prompted to generate structured rationales to explicitly justify their logic, utilizing collaborative critique to override hallucinations. We evaluate our approach on public benchmarks. Preliminary experiments guided the selection of the optimal VLM backbone, while extensive main and ablation studies suggest that scaling to a seven-agent pool with rationale-driven debate synthesizes higher-order reasoning. Experimental results show that MAD-HAR significantly improves macro-F1, while maximizing consensus and yielding consistent net error rectification.
Generative AI (GenAI) is currently being utilized in many tasks to improve their quality. Because GenAI tools are highly qualified in text-based applications, they have the potential to automate tasks across the software engineering lifecycle. In this study, we empirically investigate ChatGPT’s ability to perform code refactoring tasks. Considering five widely employed refactoring scenarios, we propose testing scenarios upon which we derive 200 test cases for five Java open-source applications. These test cases are applied using ChatGPT and NetBeans, and the results are compared and evaluated. To enable ChatGPT to perform refactoring, we follow prompt engineering approaches to design effective prompts. The results show that ChatGPT has generated refactored pieces of code, which have been successfully compiled for 88% of the test cases. However, it has correctly performed the intended refactoring for only 29% of the cases, compared to NetBeans, which has achieved a 67% correctness rate. These findings indicate that although ChatGPT has some potential to contribute to code refactoring tasks, it is not yet ready to be used as a fully automated refactoring tool for large-scale real-world applications. Its outputs still require human oversight to improve the refactored code’s correctness.
Document-level (Doc-level) event argument extraction (EAE) needs to deal with longer text inputs and complex semantic relationships than sentence-level, making it a challenging information extraction task. Extracting event arguments from an entire document primarily faces two critical issues: (i) how to handle the long-distance dependency between trigger and role arguments and (ii) how to extract key event contextual information. We propose a two-stream modeling framework using contextual clues and abstract meaning representation (AMR) parsing (TSCA). TSCA employs two-stream encoding to semantically model the document from event-critical context and event-semantic structure two perspectives. This approach leverages both contextual clues and semantic structure information to better mitigate the two issues. We incorporate AMR to assist in the semantic understanding of complex event structures and effectively capture long-distance dependencies. Additionally, we introduce a span indicator based on triggers to adaptively merge the two-stream information, enhancing the capture of semantic relevance between triggers and candidate arguments. We validated the effectiveness of our method on the public datasets RAMs and Wikievents, where TSCA achieved the best scores in various subtasks, surpassing state-of-the-art models by 3.02 F1 and 1.01 F1, respectively.
Random testing (RT) is an essential software testing strategy that selects test cases based on a random distribution. Adaptive RT (ART) encompasses a group of test-case selection/generation methods that, based on the insight that test cases that are widely spread are more prone to find nonpoint failure-causing regions, attempts to improve on the failure-finding ability of traditional RT. Two common approaches to implementing ART are distance-based ART and partition-based ART. A well-studied distance-based ART method is fixed size candidate set ART (FSCS-ART), which has been found to have superior effectiveness to RT in terms of failure-finding ability, but can incur high computational overheads. ART by bisection (ART-B) is a partition-based ART method that improves on the time efficiency of FSCS-ART. A limitation of ART-B, however, is its potential for selecting very close test cases, which reduces its failure-finding effectiveness. Adoption of an orthogonal recursive bisection (ORB) strategy in a partition-based ART method, ART by ORB (ART-ORB) has shown improvements in both failure-finding effectiveness and computational overheads; however, it also has the potential for selecting very close test cases. The present paper proposes ART-ORB with imaginary offsets (ART-ORBIO), an innovative ART method that combines the advantages of FSCS-ART, ART-B, and ART-ORB. ART-ORBIO applies a repeating input domain geometric bisection, integrated with a novel concept of imaginary offsets, to enhance the spread of test cases, and minimize the overheads of computation. Our empirical results show that ART-ORBIO provides better failure-finding effectiveness than RT and comparable effectiveness to FSCS-ART. ART-ORBIO also has significantly improved efficiency, although not comparable to ART-ORB and is more useful particularly in input domains with larger dimensions.
Model-driven engineering (MDE) has emerged as a foundational paradigm for the development of embedded systems in domains such as automotive, where stringent correctness and structural consistency must be maintained throughout the engineering process. However, traditional modeling practices often suffer from fragmented toolchains, siloed notations, and limited support for real-time collaboration across distributed stakeholders. This paper presents an artificial intelligence (AI)-augmented and role-aware collaborative framework that enables blended modeling through seamless synchronization of textual and tree-based/graphical views, underpinned by a robust service-oriented architecture. Rooted in automotive system engineering standards, specifically electronics architecture and software technology-architecture description language (EAST-ADL), and instantiated via an industrial-grade domain-specific modeling language (DSML) use case, the proposed framework facilitates stakeholder-specific modeling via both textual and tree-based/graphical editors. At the core of this infrastructure lies a real-time coordination engine that governs session orchestration, conflict resolution, and model merging, ensuring semantic consistency across heterogeneous representations. A key innovation is the integration of an explainable AI (XAI) based log analysis engine, which leverages large language models (LLMs) to generate human-readable summaries of user activity, model evolution, and design rationale. This not only enhances transparency and auditability but also introduces explainability as a first-class concern in collaborative MDE workflows. Empirical validation confirms the framework's effectiveness in enabling synchronized, traceable, and fault-tolerant co-modeling for automotive embedded systems. By fusing collaborative computing, XAI, and domain-specific rigor, this research advances the state of the art in intelligent, scalable, and human-centered system modeling environments.
Requirement prioritization (RP) is considered one of the main activities of requirement engineering (RE). The stakeholders' feedback and viewpoints in a project form the basis of RP. Two types of stakeholders can be included in a software project, one of which is the system architect and the other is the system users. After evaluating the organization and determining the non-functional requirements (NFRs), the system architects identify the technical relationships among these NFRs, while system users contribute usage-driven viewpoints that reflect their operational experiences. Stakeholders possess varying levels of organizational power, which is determined by consolidating a defined set of parameters-namely influence, project involvement, and contribution-through an expert-driven assessment. While this study focuses on these core parameters, the framework allows the incorporation of additional organizational indicators in future applications if required. This composite metric was then quantified using the $100 method to assign a final power value to each stakeholder group. An NFR's prioritization depends on both types of stakeholders. This study aimed to prioritize NFRs by collecting and extracting data from these two categories of stakeholders, leading to a heuristic smooth average approach (SAA). This approach increased the prioritization speed compared to the genetic algorithm (GA) and analytic hierarchy process (AHP) by 19% and 41%, respectively. The ease of use in prioritization also improved by 22.5% and 65%, respectively, compared to the other two approaches.
Currently, with the development of the financial market and the increase of personal credit products, the phenomenon of personal loan default has also attracted widespread attention. How to make effective predictions of loan default has become a research hotspot and difficulty in this field. This article addresses the problems in existing loan default prediction methods. Based on the domain knowledge of consumer loan defaults and machine learning theory, it is proposed to use a hybrid sampling method based on Mahalanobis distance SMOTEENN to balance the dataset. The sampling method is further extended from the original credit card fraud research to the study of loan default prediction. The balanced dataset is used to integrate categorical boosting (CatBoost)-GRU model with Bagging for consumer loan default prediction. First, the loan dataset of Lending Club platform is selected for default prediction experiments. Then, to further validate the effectiveness of the method, the method is applied to the loan default dataset of Kiva for further validation. Finally, through the comparison of the experimental results, it is proved that the method based on Mahalanobis Distance SMOTEENN hybrid sampling and Bagging integrating CatBoost-GRU can be very effective in loan default prediction. The method has strong application potential and practical effects in loan default prediction research, which not only improves the accuracy and efficiency of default prediction but also provides a wide range of insights and methodological references for the solution of similar problems.
Quantum computing is one of the research areas progressing rapidly toward practical deployment, yet the engineering of scalable and reliable quantum software remains underdeveloped. Current quantum software engineering (QSE) practices are largely tools-driven and ad hoc that providing limited support for managing probabilistic execution, hybrid quantum-classical workflows, noise sensitivity, and hardware constraints. This study proposed a structured QSE lifecycle that integrates quantum-specific characteristics with disciplined software engineering practices and principles. The proposed lifecycle organizes development into six phases, encompassing quantum requirements engineering, formal modeling, architecture and circuit design, hybrid integration, noise-aware testing, and deployment with monitoring. Each phase is supported by explicit artifacts and quantitative criteria to enable systematic progression and iterative refinement. The QSE is validated through expert assessment and simulation-based experimentation using representative variational quantum algorithms under the realistic noise conditions. The results show improved fidelity convergence, reduced resource overhead, enhanced development stability (DS), and more reliable validation compared with unstructured workflows, demonstrating the value of lifecycle-driven engineering for quantum software systems.
