Computation of Meta-Learning Classifiers in Distributed Data Mining using a Novel Cognitive Memory Model

Abstract

Distributed data mining (DDM) performs partial analysis of data at distributed locations and sends a summarized version to the peer sites or a central location for further analysis. Meta-learning is a technique that generates local classifiers (concepts or models) from distributed data sets to use in producing a global classifier. This inherently distributed nature of meta-learning provides much advantage in implementing practical DDM systems. Currently machine learning techniques such as supervised neural networks, decision trees, rules and genetic algorithms are used in the meta-learning process. Inspired by the cognitive representation of human memory, this paper presents a novel mechanism known as concept-episodic associative memory with a neighborhood effect (C-EAMwNE) to compute meta-classifiers. C-EAMwNE is an enhanced version of EAMwNE model previously developed by the authors which overcomes practical limitations of other existing cognitive representations. C-EAMwNE is applied to a multi-agent DDM system with learning agents and a central administrator agent. Learning agents use C-EAMwNE to generate meta-classifiers at distributed data sites and communicate them to the central administrator agent (CAA). CAA produces a final concept description from the distributed classifiers to be used in classification tasks.