Drift-aware ML model adaptation in data streaming systems
Loading...
Date
2026-10-12
Authors
Καγιάφα, Άννα
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The increasing use of machine learning in Internet of Things (IoT) and data-streaming applications has led to predictive models operating continuously on data generated from +dynamic real-world environments. In such settings, the conditions encountered during deployment may evolve over time, causing the distribution of incoming data to differ from that observed during initial model training. These changes can affect the predictive behavior of deployed models, motivating mechanisms that can monitor evolving data streams and support model adaptation during operation.
This thesis investigates the use of Model-Agnostic Meta-Learning (MAML) as a mechanism for runtime model adaptation in such environments and examines its integration with online change-detection mechanisms. A drift-aware streaming framework is developed that combines continuous model inference, runtime monitoring, and detector-triggered model updates. The proposed approach employs a MAML-based regression model trained across multiple related tasks to obtain a shared initialization designed for subsequent adaptation using a limited amount of recently observed data. Two complementary monitoring configurations are examined: input-distribution monitoring based on Mahalanobis distance and Adaptive Windowing (ADWIN), and prediction-performance monitoring based on changes in the model’s prediction-error stream. MAML-based adaptation is evaluated alongside cumulative Linear Regression retraining and incremental Stochastic Gradient Descent (SGD) updating.
The experimental evaluation uses environmental measurements from the Intel Berkeley Research Lab sensor dataset, with individual sensor nodes defining separate learning tasks and previously unseen sensors providing chronological streams for online evaluation. The results show that MAML supports lightweight task-specific updates using limited recent observations and incurs substantially lower update costs than cumulative retraining, while achieving prediction errors that remain close to those of conventional approaches in some online settings. The experiments further show that detected changes in the input distribution do not necessarily correspond to predictive degradation or guarantee that a subsequent model update will improve performance.
Overall, the results highlight the potential of meta-learning as an efficient mechanism for runtime model adaptation, while also emphasizing the importance of the interaction between monitoring and adaptation in evolving data streams. In particular, the findings indicate that detecting changes in the observed data and adapting a predictive model in response to those changes represent closely related, but distinct, aspects of adaptive streaming systems.
Description
Keywords
Machine learning, Meta-learning, Model-agnostic meta-learning (MAML), Data drift detection, Online model adaptation, Data streams