Maximum Variance Regularization for Latent Variables Makes Autoencoder Become Better One-Class Classifier

One-class classification (OCC) is a special kind of classification which aims to train a classifier using only one class of data. It is widely required in many fields such as outlier detection and fault detection. As a non-linear version of principal component analysis (PCA), auto-encoder (AE) has good non-linear capability and is compatible with a variety of network structures. Therefore, it is widely used for OCC tasks. In order to achieve better generalization performance, regularization is often adopted during AE training. Most of the current regularization techniques focus on the constraint on parameters, such as L1 and L2 regularization, while neglecting the network's ability to extract information. Inspired by the PCA, a variance regularization method is proposed for the latent variable space, which combines the maximum variance of latent variables and L1 regularization to improve the information extraction ability and maintain good generalization performance. Its effectiveness is validated on three datasets, MNIST dataset, TE dataset, and real industrial blast furnace data with three different AE structures: AE, gated recurrent unit (GRU)-AE and convolutional AE.