Generating Datasets for Classification Task and Predicting Best Classifiers with Conditional Generative Adversarial Networks
Ilya Kachalsky, Alexey Zabashta, Andrey A. Filchenkov, Georgiy Korneev · 2019
We focus on the algorithm selection problem and closely related dataset synthesis problem. We present conditional deep convolutional generative adversarial network we call LM-GAN, generator of which is capable of synthesizing dataset for classification in the matrix form with numeric features and discriminator of which can perform the best classifier prediction for a new never seen dataset. We also suggest a technique for transforming matrices representing datasets to a canonical form. Experimental evaluation shows that the presented network working with matrices in the canonical form outperforms baseline solutions in dataset synthesis and the best classifier prediction.