Due to its solid ability to extract local spatial features from image, convolutional neural network (CNN) has been used frequently to extract spatial features from functional near infra-red spectroscopy (fNIRS) signal. To apply CNN to multi-channel fNIRS signal, the signal has to be converted to an image. However, 2-dimensional convolution is effective only when adjacent pixels share connectivity. In order to reduce the convolution of features from unrelated channels and leave correlated channels together, graph convolution network based on functional connectivity (FC) is proposed. Spatial features extracted from different types of graphs were compared using support vector machine (SVM). Feature extracted from functional connectivity-based graph showed better performance in classifying unseen data, while 2-D convolution like methods showed sign of overfitting.