Modification of Combined Unsupervised-Supervised Cascade Scheme for Small Biomedical Data Classification

The task of improving classification accuracy in the case of analysis of short datasets is an important problem, in particular, for express diagnostics in biomedicine. Hybrid methods, including ensemble methods, are gaining more and more popularity for their solution. This paper proposes a modification of the unsupervised-supervised cascade scheme, the purpose of which is to reduce the dimensionality of the extended dataset intended for analysis. The proposed modification includes the use of the clustering method for data pre-processing, as well as the parallel application of a linear classifier at the first level of the cascade. As a result of this approach, the initial inputs of the task are replaced by only one attribute, the output signal of the linear classifier, which, together with the markers belonging to each data cluster, forms a new dataset for training the final classifier on the second cascade level. The simulation was performed on a real dataset using six different linear classifiers. Through comparison, it was established both a reduction in the time resources required for training the method and an increase in the accuracy during the analysis of small sets of biomedical data.