An optimal SVM with feature selection using multi- objective PSO

AuthorsIman Behravan, Oveis Dehghantanha, Seyed Hamid Zahiri
Conference Title1st Conference on Swarm Intelligence and Evolutionary Computation
Holding Date of Conference2016-03
Event PlaceBam
PresentationSPEECH
Conference LevelNational Conferences
KeywordsMulti-objective optimization, particle swarm optimization, Pattern Recognition, Support Vector Machines

Abstract

Support vector machine is a classifier, based on the 
structured risk minimization principle. The performance of the 
SVM, depends on different parameters such as: penalty factor, 
C, and the kernel factor, σ. Also choosing an appropriate kernel 
function can improve the Recognition Score and lower the 
amount of computation. Furthermore, selecting the useful 
features among several features in dataset not only increases the 
performance of the SVM, but also reduces the computation time 
and complexity. So this is an optimization problem which can be 
solved by a heuristic algorithm. In some cases besides the 
Recognition Score, the Reliability of the classifier’s output, is 
important. So in such cases a multi-objective optimization 
algorithm is needed. In this paper we have got the MOPSO 
algorithm to optimize the parameters of the SVM, choose 
appropriate kernel function and select the best features 
simultaneously in order to optimize the Recognition Score and 
the Reliability of the SVM. Nine different datasets, from UCI 
machine learning repository, are used to evaluate the power and 
the effectiveness of the proposed method (MOPSO-SVM). The 
results of the proposed method are compared to those which are 
achieved by RBF and MLP neural networks