Graph-Embedded Multi-Layer Kernel Ridge Regression for One-Class Classification

Humans can detect outliers just by using only observations of normal samples. Similarly, one-class classification (OCC) uses only normal samples to train a classification model which can be used for outlier detection. This paper proposes a multi-layer architecture for OCC by stacking various graph-embedded kernel ridge regression (KRR)-based autoencoders in a hierarchical fashion. We formulate the autoencoders under the graph-embedding framework to exploit local and global variance criteria. The use of multiple autoencoder layers allows us to project the input features into a new feature space on which we apply a graph-embedded regression-based one-class classifier. We build the proposed hierarchical OCC architecture in a progressive manner and optimize the parameters of each of the successive layers based on closed-form solutions. The performance of the proposed method is evaluated on 21 balanced and 20 imbalanced datasets. The effectiveness of the proposed method is indicated by the experimental results over 11 existing state-of-the-art kernel-based one-class classifiers. Friedman test is also performed to verify the statistical significance of the obtained results. By using two types of graph-embedding, 4 variants of graph-embedded multi-layer KRR-based one-class classification methods are presented in this paper. All 4 variants have performed better than the existing one-class classifiers in terms of the various performance metrics. Hence, they can be a viable alternative for OCC for a wide range of one-class classification tasks. As a future extension, various other autoencoder variants can be applied within the proposed architecture to increase efficiency and performance.

Automated summary

This paper introduces a hierarchical OCC architecture using multiple graph-embedded KRR-based autoencoders to project input features into a new feature space and apply a regression-based one-class classifier. Experimental results on 21 balanced and 20 imbalanced datasets confirm the effectiveness of the proposed method over existing kernel-based classifiers.