Topology Preserving SOM with Transductive Confidence Machine

We propose a novel topology preserving self-organized map (SOM) classifier with transductive confidence machine (TPSOM-TCM). Typically, SOM acts as a dimension reduction tool for mapping training samples from a high-dimensional input space onto a neuron grid. However, current SOM-based classifiers can not provide degrees of classification reliability for new unlabeled samples so that they are difficult to be used in risk-sensitive applications where incorrect predictions may result in serious consequences. Our method extends a typical SOM classifier to allow it to supply such reliability degrees. To achieve this objective, we define a nonconformity measurement with which a randomness test can predict how nonconforming a new unlabeled sample is with respect to the training samples. In addition, we notice that the definition of nonconformity measurement is more dependent on the quality of topology preservation than that of quantization error reduction. We thus incorporate the grey relation coefficient (GRC) into the calculation of neighborhood radii to improve the topology preservation without increasing the quantization error. Our method is able to improve the time efficiency of a previous method

k

NN-TCM, when the number of samples is large. Extensive experiments on both the UCI and KDDCUP 99 data sets show the effectiveness of our method.