An Auto-Tuning Goal-Oriented Dynamic Feature Selection for Non-Stationary Message Stream Classification Using a Multi-Agent Approach

In an organization’s digital messaging environment, information is exchanged both internally and externally, resulting in a continuous evolution of the message stream at both the conceptual and feature levels. Feature-level changes occur due to the introduction of new attributes or the evolving meaning and relevance of existing features over time. Nevertheless, the prevailing approaches to addressing this issue frequently entail the initialization of a fixed set of parameters, which can be both intricate and expensive. Furthermore, maintaining these parameters in a fixed state over time may result in solutions becoming less adaptable, efficient, or effective as the message stream evolves. This paper introduces a goal-oriented, self-tuning solution for dynamic feature selection, based on approaches from the literature. The proposed solution is applied to the classification of messages within a multi-agent system. The approach employs the iStar modeling language to integrate proactive strategies and adaptive tuning plans, which are then operationalized through the multi-agent system. The effectiveness of the solution is evaluated through test cases and an experimental studies, comparing the results of the dynamic feature selection method from the literature with the variants introduced in this work. In spam detection, the proposed solution demonstrates improvements in several of the quality measures assessed. In news classification, the proposed solution was assessed, a variant was proposed and it demonstrated improved adaptability and effectiveness. This highlights the proposed solution potential for enhanced adaptability in dynamic messaging environments.