دانلود کتاب کیفیت مقرون‌به‌صرفه برای علم داده‌ی مسئولانه: به‌کارگیری انسان در چرخه (Human-in-the-Loop) جهت دستیابی به داده‌های باکیفیت

Data preparation is a prerequisite in modern applications, in particular data analysis projects, artificial intelligence, and machine learning. The quality of data is crucial for achieving the desired results; however, due to the high volumes and the diversity of data sources, the data preparation activities that are needed to reach adequate quality levels may easily become unsustainable. Moreover, some preparation efforts may become unnecessary, if they result in negligible improvements or duplicate actions already performed in other contexts. In this book, we propose a “discount” quality approach for scientific data analysis, discussing advanced techniques for pursuing responsible data science, i.e., to design sustainable data analysis pipelines based on Human-in-the-Loop (HITL) approaches for achieving quality data. The main results focus on developing task-and context-driven sustainable HITL approaches for data preparation, data quality analysis, and data visualization, based on methods and theories to reduce the annotations and the processing space and time, which starts from the estimation of the needed human computing effort evaluation and the requirements of the specific task. The presented research focuses on the following topics: – Data ecosystems: identifying reference models and architectures to specify and describe data and pipelines in a more transparent and interpretable manner, so that they can be better found and understood by third-party consumers (humans and machines, e.g., data search engines) and eventually reused, as recommended by several EU initiatives to foster both the data economy and Open Science. Data publishers must be supported with algorithms and tools that substantially reduce the effort they can afford to dedicate to a tedious and low-priority task. – Data preparation: studying all aspects related to human intervention in data prepa-ration pipelines. This includes several activities, and in particular, we focus our research on the management of scarce resources and user motivation in repetitive tasks such as labeling, data enrichment, and cleaning. Human involvement in these activities must be systematically evaluated to minimize effort while preserving the quality and significance of the outcomes. – Data quality evaluation: evaluating the suitability of data to the task at hand in different stages of the data preparation pipelines (i.e., before and after data prepro-cessing, after data analysis) to evaluate the uncertainty of the results; classify pipeline improvements, toward methods for suggesting and minimizing human efforts in data preparation and profiling; providing a data quality assurance process for scientific data characterized by imprecise values and structures. – Information Visualization and Visual Analytics: providing a flexible human-in the-loop system that, in a data preparation cycle, identifies potential performance, data quality, and critical issues, guided by the user, resolves them to generate a new model and continues this cycle until a satisfactory model is achieved. The visual analytics layer balances the trade-off between result accuracy and interactivity by minimizing latency, ensuring a smooth exploratory data analysis experience through predictive techniques for anticipating user interactions and progressive visualizations. These combined approaches enable a more efficient workflow, allowing analysts to explore data ahead of simulation completion, confirming promising runs and discarding unproductive ones, thereby conserving computa-tional resources. Predicting user interactions also facilitates optimal resource allo-cation, focusing computational power on areas where the user prioritizes accuracy for critical decision-making. The research presented in this volume is based on the results of the project “Discount quality for responsible data science: Human-in-the-Loop for quality data,” funded by the Italian Ministry of University and Research under the PRIN Programme, a collaborative project coordinated by the RAISE (Research on Advanced Information Systems Engineering) group of Politecnico di Milano, Depart-ment of Electronics, Information, and Bioengineering, together with the DatAI Lab— Data for Artificial Intelligence of University of Milano-Bicocca, the DBGroup of University of Modena and Reggio Emilia, and the group of Sapienza Università di Roma. The aim of the book is to present the main issues and discuss the challenges of developing sustainable approaches to data preparation for researchers and companies developing systems in the data-centric AI sector.