OSMAC: A Dynamic SMAC for Data Streams

Automated machine learning (autoML) methods often require multiple passes over data and are computationally intensive, rendering them unsuitable for streaming scenarios where data is continuously generated and distributions evolve over time. The few existing autoML solutions for stream learning mainly rely on random search or genetic algorithms, which struggle to maintain high performance in dynamic environments. By contrast, leading methods in batch learning such as the Sequential Model-based Algorithm Configuration (SMAC) leverage modelbased approaches, suggesting opportunities for improvement in stream settings. To address these challenges and meet the requirements of stream scenarios, we introduce OnlineSMAC, a model-based optimizer for data streams. OnlineSMAC combines Bayesian optimization with an extension of the SMAC optimizer to dynamically select optimal processing pipelines and hyperparameters. Our results show that this approach is highly competitive, achieving performance on par with state-of-the-art stream autoML methods. This highlights the promising potential of using Bayesian optimization for data streams.