AutoOC: Automated multi-objective design of deep autoencoders and one-class classifiers using grammatical evolution

Abstract

One-Class Classification (OCC) corresponds to a subclass of unsupervised Machine Learning (ML) that is valuable when labeled data is non-existent. In this paper, we present AutoOC, a computationally efficient Grammatical Evolution (GE) approach that automatically searches for OCC models. AutoOC assumes a multi-objective optimization, aiming to increase the OCC predictive performance while reducing the ML training time. AutoOC also includes two execution speedup mechanisms, a periodic training sampling, and a multi-core fitness evaluation. In particular, we study two AutoOC variants: a pure Neuroevolution (NE) setup that optimizes two types of deep learning models, namely dense Autoencoder (AE) and Variational Autoencoder (VAE); and a general Automated Machine Learning (AutoML) ALL setup that considers five distinct OCC base learners, specifically Isolation Forest (IF), Local Outlier Factor (LOF), One-Class SVM (OC-SVM), AE and VAE. Several experiments were conducted, using eight public OpenML datasets and two validation scenarios (unsupervised and supervised). The results show that AutoOC requires a reasonable amount of execution time and tends to obtain lightweight OCC models. Moreover, AutoOC provides quality predictive results, outperforming a baseline IF for all analyzed datasets and surpassing the best supervised OpenML human modeling for two datasets.