2.3.2. Support Vector Machine (SVM) and Principal Component Analysis (PCA)

The SVM classifier also has 900 inputs and three outputs. It is designed using the Matlab Classification Learner App. Error-correcting output codes (ECOC) [33] are used to train the classifier which works by solving for a hyperplane that separates two class data with maximal margin [34]. The support vectors are the points which lie near the separating hyperplane. The SVM is trained for a 3-class problem on a one vs all approach. Since the original training data is not linearly separable, four different kernel functions K(x_i,x) (linear (8), quadratic (9), cubic (10), and Gaussian (11)) are applied to the classifier. These transform the original input space into vectors of a highly dimensional feature space for the SVM to classify. The general structure of the SVM is shown in Figure 8.

SVM architecture.

Using 900 predictors in the input space can impair computational time. Hence, PCA [35] is used to determine the principal components for feature optimization [36,37]. This singular value decomposition reduces the data dimensionality and projects it to a lower dimensional environment for the SVM. This naturally comes at the expense of accuracy. Hence, it is important to compare the SVM + PCA classifier accuracy with the SVM only classifier.