Search

The Support Vector Machine (SVM) is a supervised learning algorithm to analyze data and recognize patterns. The standard SVM suffers from some limitations in nonlinear classification problems. To tackle these limitations, the nonlinear form of the SVM poses a modified machine based on the kernel functions or other nonlinear feature mappings obviating the mentioned imperfection. However, choosing an efficient kernel or feature mapping function is strongly dependent on data structure. Thus, a flexible feature mapping can be confidently applied in different types of data structures without challenging a kernel selection and its tuning. This paper introduces a new flexible feature mapping approach based on the Dirichlet distribution in order to develop an efficient SVM for nonlinear data structures. To determine the parameters of the Dirichlet mapping, a tuning technique is employed based on the maximum likelihood estimation and Newton’s optimization method. The numerical results illustrate the superiority of the proposed machine in terms of the accuracy and relative error rate measures in comparison to the traditional ones. [ABSTRACT FROM AUTHOR]

Support vector machine (SVM) is a popular supervised learning algorithm achieved much success in classification problems in both single output and multi-output cases. Despite this fact, SVM has some limitations, for example, in terms of the accuracy when data are not linearly separable. One of the well-known tricks for tackling this obstacle is using kernel functions in the base model of SVM which map all data into a higher or infinite dimension space called feature space. We propose polar coordinate system as an efficient technique for mapping data and improving the accuracy of SVM in some datasets. Hence, the current research is assigned to extending SVM model to polar one for single and multi-output cases. Also, the proposed SVM may be combined with kernel functions to outperform traditional non-linear SVM. The performance of the proposed model is compared with traditional one to affirm the superiority of polar SVM. [ABSTRACT FROM AUTHOR]

Highlights: [•] This paper addresses a new clustering-classification method for classifying the severity of road accidents in Iran. [•] k-Means and some classification methods are used for running numerical experiments. [•] Some activation functions have been used for training the neural networks. The results of them are compared together. [•] Also, the accuracy of classification is compared by using different approaches in k-mean. [•] Silhouette technique was used to confirm the validity of the clusters. [Copyright &y& Elsevier]