Graph Convolutional Network Based Generative Adversarial Networks for the Algorithm Selection Problem in Classification

Abstract

In this work, we address the algorithm selection problem for classification via meta-learning and generative adversarial networks. We focus on the dataset representation question. The matrix representation of classification dataset is not sensitive to swapping any two rows or any two columns. We suggest a special method to reduce a dataset to a unified form. This allows to apply generative adversarial networks to classification dataset generation. In this setting, a generator generates new classification datasets in a matrix form, while a conditional discriminator is trying to predict for a dataset and an algorithm if the dataset is real and the algorithm would show the best performance on this dataset. We also suggest a graph convolutional network as a discriminator that is capable to work with such forms, which encode a dataset as a weighted graph with nodes representing objects.