Online machine learning in streaming applications

Stavros Kontopoulos and Debasish Ghosh explore online machine learning algorithm choices for streaming applications, especially those with resource-constrained use cases like IoT and personalization. They dive into Hoeffding Adaptive Trees, classic sketch data structures, and drift detection algorithms from implementation to production deployment, describing the pros and cons of each of them.

Applications such as smart homes, smart monitoring of industrial environments, augmented reality in retail, and autoconnected cars are driving a new era in online ML, where ML algorithms have been moved to the edge instead of the cloud. These applications are constrained in terms of resources like power, CPU, memory, etc. and responsiveness. Data flows in the system and the application needs to interact with the surrounding environment in a given time window. Stavros Kontopoulos and Debasish Ghosh explore the foundations of the algorithmic aspects (Hoeffding Adaptive Trees, classic sketch data structures, and drift detection algorithms) of these applications and dive into the details of how they can be implemented and deployed efficiently in production. They evaluate production concerns like performance (latency, memory footprint, etc.), techniques for updating models being served in a running pipeline and future trends like feature space representation and sampling, and tools to use for the actual implementation and deployment of these algorithms. The concepts they detail can also be applied in a cloud setting, so many of the practical aspects they cover are universal and will benefit any practitioner of ML. The main focus is cutting-edge applications and technologies, and you don’t want to miss a glance in the future.