Adaptive AutoML Techniques for Real-Time Predictive Modeling in High-Velocity Streaming Data: A Comprehensive Review
Abstract
The exponential growth of Internet of Things (IoT) sensors, financial transaction networks, digital health monitors, and web telemetries has ushered in an era of continuous, high-velocity streaming data. Modern machine learning applications increasingly demand real-time predictive modeling capable of delivering low-latency inferences while operating under tight memory and computational constraints. However, traditional Automated Machine Learning (AutoML) frameworks—designed predominantly for static, offline datasets—fail in streaming environments due to non-stationary data distributions (concept drift), unbounded data volumes, and catastrophic forgetting. This paper provides a comprehensive, state-of-the-art review of Adaptive AutoML techniques tailored specifically for high-velocity streaming environments. We systematically explore the architectural paradigms that enable automated pipeline search, dynamic feature selection, adaptive hyperparameter optimization, and incremental model selection in real-time. Key drift detection algorithms, including ADWIN, Page-Hinkley, and DDM, are analyzed alongside incremental learning algorithms such as Very Fast Decision Trees (VFDT), Hoeffding Adaptive Trees (HAT), and streaming ensemble methods. Furthermore, we evaluate resource-aware meta-learning mechanisms designed to maintain sub-millisecond inference latencies. Through a detailed comparative taxonomy, experimental performance evaluation, and critical analysis of existing frameworks (e.g., StreamML, River, and Scikit-Multiflow), we highlight the critical trade-offs between predictive accuracy, memory footprint, adaptation speed, and computational throughput. Finally, we identify prominent research gaps and chart future directions, including federated streaming AutoML, edge AI deployment, and explainable adaptive streaming models.
KEYWORDS: Automated Machine Learning (AutoML); Streaming Data; Real-Time Predictive Modeling; Concept Drift Adaptation; Incremental Learning; Meta-Learning; Resource-Constrained AI.
Full Text:
PDF 69-86Refbacks
- There are currently no refbacks.