A Moment-Based Classification Algorithm and Its Comparison to the K-Nearest Neighbours for 2-Dimensional Data

The problem of probability density function reconstruction from statistical moments has been studied since 19th century. However, there has not been an attempt toapply the moments directly to classification problems. In thismanuscript, we proposed a classification method of orthogonalmoments for classification of two-dimensional data. We comparedthe proposed moment-based classifier with a well-known non-parametric classification algorithm: k-nearest neighbours. Theproposed method as well as k-nearest neighbours are bothprobability density function reconstruction methods, however, theproposed method does not require storing the whole data setthus making it computationally faster. In terms of accuracy bothmethods were tested on a synthetically generated data and thefamous Fisher's Iris data set that we reduced to two attributes. The results show that the moment-based classification achievessimilar classification accuracy on the Fisher's Iris data and forcertain parameters outperforms k-nearest neighbours classifier onthe highly nonlinear synthetically generated data set.