Shallow and Ensemble Deep Randomized Neural Network for Anomaly Detection

Abstract

Anomaly detection, or one-class classification (OCC), plays a vital role in real-world applications. Traditional support vector machine-based OCC models often face challenges with large-scale datasets and are sensitive to the choice of kernel functions. To overcome these limitations and enhance the generalization of OCC, we propose the one-class random vector functional link (OC-RVFL) network, which fuses both linear and nonlinear patterns through a combination of original and randomized features. While OC-RVFL efficiently computes output weights by solving linear equations, its single hidden layer restricts its ability to capture complex patterns. To address this, we introduce the one-class ensemble deep RVFL (OC-edRVFL), a fusion of deep learning and ensemble learning principles, with OC-RVFL as the base model in each layer. OC-edRVFL, the novel fusion model for OCC provides enhanced stability, robustness, and generalization compared to OC-RVFL. The proposed models employ a closed-form solution for output weight computation, reducing training time. We also derive an upper bound on the generalization error for these models. Experiments on artificial, UCI, NDC, and MNIST datasets demonstrate that OC-edRVFL outperforms baseline models, showcasing its superior performance even with training sizes of up to 5 million samples.

Keywords: One-Class Classification, deep learning, Ensemble Learning, Multiple output layers, Random Vector Functional Link (RVFL) Network, Ensemble Deep RVFL.