Fast classification in online and nearline modes in conditions of overlapping classes

The subject of the research is the process of fast classification under overlapping classes conditions. The aim is to develop an approach to fast classification that combines online and nearline modes to improve classification accuracy in overlapping classes conditions. Objectives: development of a multilayer neural network with kernel bell-shaped activation functions for online mode operation; development of an adaptive neuro-fuzzy system for nearline mode operation. The scientific novelty of the obtained results is as follows. For the first time, an approach to fast classification is proposed that combines online and nearline modes to solve problems of refining boundaries between classes, adapting to changes in input data distribution, eliminating class imbalance, and removing noise points. A multilayer neural network with kernel bell-shaped activation functions is proposed for solving the classification problem under overlapping classes conditions in online mode. The network uses a recurrent learning algorithm, which allows avoiding the backpropagation procedure and the «vanishing gradient» effect. In practical terms, the network provides system adaptation to complex boundaries between classes and creates conditions for adapting to changes in input data distribution. An adaptive neuro-fuzzy system for classifying data in matrix form has been developed, which uses hybrid combined learning. Learning is implemented in sequential online or nearline mode as new data arrives. Experimental verification of the proposed approach showed the possibility of maintaining classification accuracy under overlapping classes conditions and changes in input data distribution.