Regularized Tensor Learning with Adaptive One-Class Support Vector Machines

Abstract

The extraction of useful information from multi-sensors data requires fairly involved methodologies and algorithms. We propose an $L_1$ regularized tensor decomposition to decrease learning sensitivities, coupled with an adaptive one-class support vector machine (OCSVM) for anomaly detection purposes. This new framework yields sparse and smooth representations of the desired outcomes. An automatic parameter selection method based on the euclidean metric is also proposed to adaptively tune the kernel parameter inherent in OCSVM. These positive characteristics of tensor analysis allow us to fuse data from multiple sensors and further analyze them at the same time at which informative features are being extracted. This work is challenging because it is cross disciplinary; and thus it requires coherency to the specific domain applications fundamentals (such as structural health monitoring), on the one hand, and its diversity on machine learning techniques on the other. Compared to the state-of-the-art approaches for learning tensor and anomaly detection, our proposed methods work well on experiments and show better performance in terms of decomposition quality and stability of the extracted features.