Swarm Search Methods in Weka for Data Mining

Abstract

Building a good prediction from high-dimensional data model in data mining is a challenging endeavor. One key step in data pre-processing is feature selection (FS) which is about finding the right feature subset for effective supervised learning. FS has two parts: feature evaluators and search methods to find the appropriate features in the search space. In this paper we introduce a collection of search methods that implement metaheuristics search which is also known as swarm search (SS). SS has the advantage over conventional search such as local search, that SS has the facility to explore global optima by a group of autonomous search agents. We have recently added nine new methods to the Weka machine learning workbench. The objective of these nine swarm search methods is to supplement the existing search methods in Weka for providing efficient and effective FS in data mining. We have carried out two experiments using synthetic data and medical data. The results show that in general SS has certain advantages over the conventional search methods. The SS methods can be found in the Weka Package Manager as open source code. Researchers and Weka users are encouraged to enhance data mining performance using these free swarm search programs.