Minimum Variance-Embedded Multi-layer Kernel Ridge Regression for One-class Classification

Abstract

In this paper, a Multi-layer architecture is proposed by stacking minimum Variance-Embedded Kernel Ridge Regression (KRR) based Auto-Encoder in a hierarchical fashion for One-class Classification, and is referred toVMKOC. Two types of Auto-Encoders are employed for this purpose. One is vanilla Auto-Encoder and other is Variance-Embedded Auto-Encoder. The first one minimizes only reconstruction error and the latter one minimizes the intra-class variance and reconstruction error, simultaneously within the multi-layer architecture. These Auto-Encoders are employed as multiple layers to project the input features into new feature space, and the obtained projected features are passed to the last layer ofVMKOC. The last layer of VMKOC is constructed by KRR-based one class classifier. The extensive experiments are conducted on 17 benchmark datasets to verify the effectiveness ofVMKOC over 11 existing state-of-the-art kernel-based one-class classifiers. The statistical significance of the obtained outcomes is also verified by employing a Friedman test on the obtained results.