Abstract:We introduce an algorithm for one-class classification based on binary classification of the target class against synthetic samples. We use a process inspired by Generative Adversarial Networks (GANs) in order to both acquire synthetic samples and to build the one-class classifier. The first objective is achieved by leading the generator’s output into close vicinities of the target class region. For the second objective, we obtain a one-class classifier by generating an ensemble of discriminators obtained from the GAN’s training process. Our approach is tested on publicly available datasets producing promising results when compared to other methods. Forthcoming related work concerns the extension of our GAN based algorithm to non-image data, and applying it, along with other deep learning techniques, to predicting risk of dementia using routine primary care records (CPRD), in a project between Division of Population Health, Health Services Research and Primary Care at University of Manchester, and the Data Science & Soft Computing Lab in London.

Publication: Mihai Ermaliuc, Daniel Stamate, George Magoulas, Ida Pu;
Creating Ensembles of Generative Adversarial Network Discriminators for One-class Classification, Proceedings of 22nd International Conference of Engineering Applications of Neural Networks, EANN 2021, Springer, June 2021.

https://link.springer.com/chapter/10.1007/978-3-030-80568-5_2

Preprint https://www.doc.gold.ac.uk/~mas01ds/dssc/pub/GANOCC_EANN21.pdf

Mini bio: Mihai Ermaliuc develops his PhD research in the broad area of Neural Networks and Deep Learning, focusing on Generative Adversarial Networks (GANs) algorithms, and he plans to apply these cutting-edge neural network techniques in mental health prediction modelling, in particular in dementia prediction, as a next step in his PhD. Mihai conducts his research in the Data Science & Soft Computing Lab in the Department of Computing, Goldsmiths, University of London. He has a background in Computer Science (BSc, Bucharest) and Data Science (MSc, London), and is currently a lead data scientist professional in industry collaborating in this role also with academics for research, while doing his PhD part time. In his talk, Mihai will introduce a new technique based on GANs, that is able to achieve high performance in the one-class classification problem. His work was recently nominated for the best PhD student paper award of the International Conference of Engineering Applications of Neural Networks, EANN 2021.