Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 4: Data quality process framework
Introduction: Artificial intelligence (AI)-related products, systems or solutions have developed quickly in recent years. One of the common characteristics of an AI system, especially for systems using supervised machine learning (ML), is whether the AI system can be trained on a dataset before deployment or trained dynamically as the system is used. Data have been recognized as one of the most important aspects of ML-based AI systems. For all supervised, semi-supervised, unsupervised and reinforcement learning approaches, data quality can be a primary concern in creating and using data for training and evaluating ML systems. It has been shown that with more accurate and richer data, the results of analytics and ML can be more useful and reliable. In addition, for the development of supervised learning-based AI systems, a large number of task-specific labelled training data is needed. This makes accurately labelled data one of the most important resources in the AI industry. Nowadays, there is a verified market of industrial services and tools for training data labelling. This market is now reaching a level of maturity that justifies the development of International Standards for the benefit of providers and users of these services and tools to ensure high-quality labelled data. This document describes the implementation of a standardized common procedure of data processing with regard to data quality for analytics and ML. Clause 5 describes principles about data quality process and Clause
Scope: This document establishes general common organizational approaches, regardless of the type, size or nature of the applying organization, to ensure data quality for training and evaluation in analytics and machine learning (ML). It includes guidance on the data quality process for: — supervised ML with regard to the labelling of data used for training ML systems, including common organizational approaches for training data labelling; — unsupervised ML; — semi-supervised ML; — reinforcement learning; — analytics. This document is applicable to training and evaluation data that come from different sources, including data acquisition and data composition, data preparation, data labelling, evaluation and data use. This document does not define specific services, platforms or tools.