Data-Centric MLOps Frameworks for Large-Scale Intelligent Service Environments
Keywords:
Dataset Quality Management; Ubiquitous Computing; Active Learning; Data Validation and Monitoring; Edge–Cloud MLOps; Concept Drift; Large-Scale AI SystemsAbstract
The large-scale growth in the ubiquitous computing and large-scale intelligent service environments in the form of smart cities, industrial IoT, intelligent transportation, and context-aware cyber-physical systems have revealed inherent vulnerabilities in model-centric machine learning deployment practises, especially in the face of noisy, heterogeneous and constantly changing streams of data. This review then responds by providing an overview of data-centric MLOps frameworks based on data quality management and continuous dataset enhancement as firstclass machine learning life cycle processes. The article systematically reviews the main data issues of intelligent environments such as data noise, redundancy, sparse labels, biasing and concept drift and reviews the existing state-of-the-art data cleaning and deduplication approaches, active learning, weak supervision, and automated data validation within large-scale pipelines of MLOps. Moreover, architectural designs and deployment models of embedding data-driven workflows in edge-cloud systems are presented and illustrative cases of application, as well as metrics used to evaluate performance indicators of system resilience and changes that exhibit the effects that dataset quality has on both system robustness and system adaptability. Through synthesis and purposeful gaps of existing advancements, the research can offer a concise view of how data-centric MLOps may facilitate reliable, scalable and sustainable intelligent services in next generation ubiquitous computing operational settings

