Answers are grounded in this paper and cite their sources.
Data is increasingly generated as continuous, high-speed streams produced by sensors, transactional systems, social platforms and networked devices operating in dynamic, non-stationary environments. Traditional batch learning, which assumes finite training sets and static, stationary distributions, is not appropriate for this setting. This article reviews the foundations and recent advances of real-time data mining, organised around three pillars: (i) clustering evolving data streams, from sequential k-means to micro-cluster-based frameworks such as CluStream and ClusTree, and methods for monitoring cluster evolution over time; (ii) predictive learning from streams, centred on Hoeffding-bound decision trees (the Very Fast Decision Tree family), their extensions to regression, rules and ensembles, and a generic memory/learning/loss-estimation/change-detection schema that organises how such algorithms detect and adapt to concept drift, evaluated through the prequential methodology for unbiased assessment; and (iii) a real-world case study in predictive maintenance of railway assets, where deep sequence models are used for real-time failure detection and explanation. We close with lessons learned and a set of open research issues, including automated machine learning for streaming algorithms, learning from structured and networked data, and semi-supervised and active learning under drift. The overarching message is that, in time-evolving domains, learning is not a one-shot event but a continuous process that must itself be monitored, diagnosed and adapted.
No takes yet. Share an insight, caveat, or question.
João Gama (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: