Support vector machines in multisource classification

The multiclass classification problem considered here was reduced to multiple margin-based binary problems, and an extension of the radial basis function kernel for multisource data is proposed, which concentrates on local distance between features from each data source.

Abstract

The use of Support Vector Machines (SVMs) for classification of multisource data is investigated. SVMs have been shown to have difficulties in classifying multiclass data. To over- come that,the multiclass classification problem considered here was reduced to multiple margin-based binary problems. Several possibilities of binary problems were investigated,including one- against-all,all-pairs and nearly random decompositions of the multiclass problem. To combine the outputs from the binary problems three approaches were tested: a) voting schemes,b) two loss functions,and c) decoding function based on condi- tional probability estimation. An extension of the radial basis function kernel for multisource data is also proposed. The kernel concentrates on local distance between features from each data source. The experimental results show the proposed approach to be appropriate for multisource data classification. I.I NTRODUCTION