Adaptive data processing using a machine learning mechanism

Abstract

The relevance of this topic stems from the rapid growth in the number of heterogeneous and unstable data sources in modern information systems. Under such conditions, traditional approaches to organizing information retrieval no longer ensure sufficient efficiency, as they do not account for changes in the availability of sources, their response times, and the quality of the results obtained. The problem becomes particularly relevant in event-driven architectures, for which there are currently no universal mechanisms for adaptive control of the search process and selection of data sources. The increasing volume of publications requires innovative solutions that directly impact the quality of the results obtained and the overall efficiency of computational resource utilization. The aim of this article is to develop and implement reinforcement learning algorithms in event-driven systems to solve problems of adaptive evaluation and selection of data sources. To achieve this goal, an adaptive selection mechanism based on the Soft Actor-Critic (SAC) [9], [13] algorithm has been developed. An approach is proposed for forming a source state vector and a reward function based on behavioral and content metrics. This allows the system to dynamically adapt its search strategy depending on the characteristics of the query and the effectiveness of the sources in real time. Unlike static meta-search strategies, the developed system is capable of independently adapting to environmental changes during operation. A simulation environment was created, the results of which confirmed the unquestionable advantage of the adaptive approach: balance is achieved between result quality (increased relevance) and costs (reduced processing time and elimination of duplicates).

Keywords: adaptive source selection, event-driven systems, machine learning, reinforcement learning, Soft Actor-Critic, asynchronous data processing, ranking, deduplication, distributed search systems, reward function, state vector, action vector, scientific information aggregation, semantic search, relevance, dynamic control, query optimization, metasearch systems.