GAN-SSL Classification for Identification Expertise in Chemistry

Abstract

In this work we investigate the generative adversarial nets for classification problem of identification expertise in Chemistry. The identification expertise problem is challenging for classification because of complex structure of classes, outliers and cross-classes. The generative-adversarial nest for semisupervised learning (GAN-SSL) is proposed for complex classification problem. The training samples are partially labeled for the semi-supervised tasks. Two groups of experiments were carried out. The first group of experiments for the model dataset that consist of classes of points normally distributed about vertices an eightdimensional hypercube. The second groups of experiments for the petrol identification expertise dataset we get from laboratory of petrol quality. The experiments with good model examples get good quality more than 99%. The classification model for petrol identification expertise was created and has 93% quality but convergences training much worse. In this work we use GAN-SSL classification on petrol identification expertise example, but this classification model can be used for diesel fuel, household chemicals items, different oils and for various other objects.