Classification Ability of Self Organizing Maps in Comparison with Other Classification Methods

摘要

In this study, performance of different classification methods based on self organizing maps (SOMs), including counter propagation network (CPN), supervised kohonen networks (SKN) and XY-fused network (XYF), were compared to linear discriminant analysis (LDA), learning vector quantization (LVQ) and support vector machine (SVM). Performance of classification was statistically investigated using both simulated and real data sets according to percent of correct classified samples (non-error rate) in the test set. Effect of selection of calibration samples on model stability and performance was investigated. There were several adjustable parameters in each modeling techniques (except LDA) which were optimized for better comparison. Each simulated dataset regenerated 50 times and performance of classification was computed for each method to obtain a population of results. Obtained results showed that the distribution and structure of the samples in data space is an important factor influences on the relative performance of classification methods. CPN and SVM performed better in the cases with nonlinear discriminant boundary but for overlapped classes with normal distribution, performance of LDA was slightly better than other methods. In addition, CPN showed a comparable performance and stability in comparison with SVM which known as a powerful classification method.