The Smaller, the Better

Abstract

Support vector machine (SVM) is a supervised classifier which has been applied for solving a wide range of pattern recognition problems. However, training of SVMs may easily become a bottleneck, because of its time and memory requirements. Enduring this issue is a vital research topic, especially in the era of big data. In this abstract, we present our adaptive memetic algorithm for selection of refined (significantly smaller) SVM training sets. The algorithm - being a hybrid of an adaptive genetic algorithm and some refinement procedures - exploits the knowledge about the training set vectors extracted before the evolution, and attained dynamically during the search. The results obtained for several real-life, benchmark, and artificial datasets showed that our approach outperforms the other state-of-the-art techniques, and is able to extract very high-quality SVM training sets.