Comparison of Kernels in Online SVM Classification of fNIRS Data

• Published in: 2018 18th International Conference on Control, Automation and Systems (ICCAS) pp. 1152 - 1157
• Main Authors: Huang, Ruisen, Kavichai, Eakdanai, Hong, Keum-Shik
• Format: Conference Proceeding
• Language: English
• Published: Institute of Control, Robotics and Systems - ICROS 01.10.2018
• Subjects: Feature extraction; functional near-infrared spectroscopy; Kernel; support vector machine; Support vector machines; Task analysis; Training; Training data

The purpose of this study is to reduce the online data processing time while maintaining good classification accuracy. For this purpose, kernel techniques are important means to enable a support vector machine (SVM) classifier to distinguish the non-linearly separable data set. This study investigated the performance of Gaussian kernel-based SVM and polynomial kernel-based SVM for processing functional near-infrared spectroscopy (fNIRS) data. Eight subjects participated in the experiment and performed 5 trials of mental arithmetic tasks. The brain signals were acquired using an fNIRS system from the prefrontal cortex. Five distinctive features (the slope, mean, variance, maximum and minimum of signal) were extracted and 3,000 samples of signals were used to train the SVM classifier. 16 \times 8 classifiers were tested with a Gaussian kernel using 16 values of box constraint ( C) and 8 values of standard deviation ( \sigma ), while 16 \times 9 classifiers were tested with a polynomial kernel using 16 values of C and 9 values of the kernel order ( p). The performance of the classifiers with different kernels and various configurations showed that, the Gaussian kernel (74.63% on average) outperformed the polynomial kernel (73.71% on average), but the latter was found to be more stable. The computation time analysis of both kernels shows great variation and it can be concluded that the polynomial kernel-based SVM is more robust for online systems due to the absence of singularity problem in numeric approximation.