Machine learning applications for event routing in streaming systems

Abstract

The paper provides a broad overview and classification of machine learning methods used to optimize routing in distributed streaming architectures. The aim of the study is to provide a detailed analysis of existing approaches: from classical reinforcement learning algorithms to modern deep neural networks, with an assessment of their potential in various operational scenarios and identification of key limitations. The methodological basis was a systematic review of publications dealing with intelligent routing, real-time data processing, and integration of ML solutions into system pipelines. Three main classes of algorithms were identified and considered: reinforcement learning methods (including DQN and actor-critic), deep networks (CNN, RNN and their hybrids), as well as ensemble and evolutionary techniques. The advantages and disadvantages of each class are analyzed in terms of key criteria — response time to flow changes, scalability in the number of nodes, and the ability to dynamically adapt. Special attention was paid to hybrid strategies that combine several models to increase the reliability and accuracy of recommendations on event transmission routes. In conclusion, the main conclusions about the current state of research are formulated and promising areas are outlined: the development of more robust architectures with explicable decision-making logic, as well as the integration of graph neural networks for modeling complex topologies of distributed systems. The presented results will be useful for engineers developing streaming platforms, big data analysis specialists, and research groups working on information channel optimization tasks.