Consensual based classification as emergent decisions in a complex system

In massive multi-agent systems that are used to model some complex systems, emergence is a key feature that allows to model high-level states of such systems. According to this perspective, the work we introduce in this paper entails the handling of emergence in massive multi-classifiers that we consider as complex systems. We aim to build a collaborative system for supervised data classification that we expect to provide better performance, compared to conventional classifiers. Modeled as a multi-agent system, the massive multi-classifier is composed of a high number of agents that are interconnected according to a given neighborhood. Each agent plays the role of a weak classifier. At the micro-level, the elementary interaction between agents consists of combining their respective classification results. Every agent, according to the majority vote rule, combines its result with those of its neighbors by taking into account their respective performances. This process is iterated continuously in a cyclic manner within the neighborhood of each agent. Therefore, a complex dynamic will be created within the system. After a certain time, this complex dynamic stabilizes, allowing the exhibition of an emergent structure that will be observed at the macro-level and is considered as a consensual class prediction for the data we want to classify. Obtained experimental results and the comparison with conventional classifiers show the potential of the approach to enhance classification and to be an alternative for classifier combination and aggregation.