
ABSTRACT Although previous studies have examined maintenance and evolution in embedded and related software contexts, cross‐domain empirical evidence centered on practitioner‐reported practices remains fragmented. This study provides an exploratory characterization of embedded‐software maintenance and evolution based on a survey of 36 practitioners from different professional roles, experience levels, and application domains. The questionnaire investigated maintenance characteristics and difficulties, reported strategies and tools, planning and rework, lifecycle impact, maintenance metrics, perceptions of new technologies, agile‐method adaptations, and expectations regarding a unified maintenance and evolution model. Data were collected online from August to October 2024. Structured responses were analyzed primarily through descriptive frequencies and percentages, complemented by qualitative coding of free‐text responses and additional comments. Lack of documentation (26 respondents), integration of new technologies (22), and hardware limitations (16) were the most frequently reported maintenance difficulties. Code review (26), code refactoring (23), and configuration management (14) were the most frequently reported maintenance strategies, while Git/GitHub was reported by 33 of the 36 participants. Twenty‐four participants reported using agile methodologies in embedded‐software projects. Only two participants reported prior use of a unified maintenance and evolution model; consequently, responses concerning its components and benefits are interpreted as expectations and perceived potential benefits rather than evidence of established model effectiveness. The findings provide practitioner‐informed descriptive evidence and identify areas for subsequent replication and empirical evaluation rather than population‐level or causal conclusions.
Software defect prediction supports targeted quality assurance by identifying modules that warrant early testing and maintenance attention. This study presents an Explainable Graph Neural Network (X-GNN) framework for class-level defect prediction, in which software versions are modeled as directed dependency graphs. Each class is represented by a hybrid feature vector combining code metrics with graph-derived structural metrics. A Graph Convolutional Network (GCN) propagates information through class dependencies, while an attention mechanism assigns nonuniform importance to neighboring classes. SHapley Additive exPlanations (SHAP) are applied only after model training to interpret the resulting predictions. The empirical study is deliberately bounded to Defects4J v3.0.1, which contains 854 active reproducible defects across 17 Java projects. Under the adopted within-project stratified train/validation/test protocol, X-GNN achieved an accuracy of 0.85, F1-score of 0.83, precision of 0.82, recall of 0.84, and ROC-AUC of 0.88, outperforming random forest (RF), logistic regression (LR), and standard GCN. On the closure project, X-GNN attained a ROC-AUC of 0.94, compared with 0.90 for standard GCN and 0.87 for RF. SHAP-based evaluation on the Lang project yielded a fidelity of 0.94, coverage of 0.91, and node importance agreement (NIA) of 0.87. The project-level F1-score comparison between X-GNN and Standard sCN produced p = 0.003 with Cohen's d = 0.75. The results provide benchmark-specific evidence that attention-enhanced graph learning improves class-level defect prediction, while post hoc attribution provides model-centric interpretive information. The empirical claims are restricted to Defects4J and do not imply cross-benchmark generalization or cross-method explanation agreement.
One of the effects of COVID-19 was that software engineers were forced to work remotely, in many cases without prior experience, which may have led to a number of challenges for projects and organizations, along with developers' well-being. Several work styles were considered after the pandemic, because not only did teleworking persist but new forms of work emerged, for which neither organizations nor employees were prepared. These new modalities have led to challenges that have not yet been sufficiently addressed in the existing literature, mainly because they are relatively recent phenomena. The aim of this paper is to analyze the risks that remote or hybrid working can pose to software development and examine them from the perspective of the company, the project itself, and employees. Once the problems have been identified, it is important to analyze them from the point of view of a software engineer's personality and assess the likelihood of these challenges arising according to the working model used. It is also essential to find strategies with which to mitigate these challenges. These are the objectives of this paper, which aims to help project managers and developers carry out hybrid or remote software development efficiently and ensure their employees' well-being. To achieve these objectives, a mixed-methods approach comprising five stages was adopted: a systematic mapping review to identify and categorize the factors affecting software engineers, a survey of 184 practitioners to validate and refine these factors, a second survey examining the perceived impact of different work modalities on productivity and associated risks, and an evaluation of our framework (main contribution of this paper) through semi-structured interviews with 10 practitioners.
Hugging Face (HF) is an increasingly popular platform for sharing and discovering machine learning (ML) and artificial intelligence (AI) models. Model cards document these models, and good model cards serve as a vital communication bridge between model developers and users. Without good model cards, it is challenging for users to understand the models and for developers to maintain them. Prior research focused on the documentation process and automatic generation of model cards, and identified the need for high-quality datasets. Understanding the characteristics of good model cards can help improve the quality of model cards, which is essential for model developers and users. To address these gaps, we study the characteristics of high-quality model cards. We collect 758 sampled models along with their model cards and adapt a prior framework to manually identify seven categories of content (i.e., categories of "What," "Why," "How," "When," "Who," "References," and "Other") and their corresponding themes (e.g., "Model Description" in the "What" category) for these model cards. Then, we conduct a survey to explore what constitutes a high-quality model card. After validation, we define a high-quality model card as one that incorporates the "Model Overview" and "Model Evaluation" content themes from the "What" category, along with the "Model Usage Guide" content theme from the "How" category, that is, a model card that provides both a brief or detailed introduction to the model and instructions for its usage. Then, we train classifiers to investigate whether high-quality model cards can be automatically identified. The random forest model achieves the best performance, with an ROC-AUC of 94.57%. We also investigate the determinants of the quality of model cards. We find that the total word count of the content of model cards, the number of tables that show the key evaluation metrics and results, external links that show the references of external resources, and code blocks that are used to show examples of how to use the model are several of the most significant features that determine the quality of model cards. Finally, we explore the relationship between model card quality and model survival, and find that models with high-quality model cards tend to remain active significantly longer than those with low-quality model cards. Our findings provide insights into high-quality model cards and practical suggestions for model developers, users, researchers, and the HF platform.
ABSTRACT Deep learning is the core technology underlying many artificial intelligence (AI) applications. Maintaining deep learning software programs presents unique challenges due to their complex, interdependent structures. Traditional change impact analysis methods, while effective for conventional software, fall short when they are applied to deep learning programs due to the intricate interdependencies within neural networks, particularly those related to the unique architectural layers and hyperparameters. This paper presents a novel approach to change impact analysis which is specifically designed for deep learning programs. We introduce a comprehensive taxonomy that systematically categorizes changes in deep learning architectures, hyperparameters, and regularization techniques. Our approach leverages this taxonomy to map changes automatically, thereby enabling structured and efficient impact analysis in complex deep learning programs. To predict the potential impact of initial changes, our framework incorporates co‐change pattern mining to identify and generate co‐change rules, capturing dependencies based on historical co‐change patterns. We also designed domain‐expert rules derived from deep learning literature and best practices. Our evaluation results demonstrate that our co‐change mining approach effectively balances recall and precision, with the integration of co‐change patterns and domain‐expert rules offering additional complementary benefits.