Anomaly detection in homogenous populations: A sparse multiple kernel-based regularization method

A problem of anomaly detection in homogenous populations consisting of linear stable systems is studied. The recently introduced sparse multiple kernel based regularization method is applied to solve the problem. A common problem with the existing regularization methods is that there lacks an efficient and systematic way to tune the involved regularization parameters. In contrast, the hyper-parameters (some of them can be interpreted as regularization parameters) involved in the proposed method are tuned in an automatic way, and in fact estimated by using the empirical Bayes method. What's more, both the parameter and hyper-parameter estimation problems can be cast as convex and sequential convex optimization problems. It is possible to derive scalable solutions to both the parameter and hyper-parameter estimation problems and thus provide a scalable solution to the anomaly detection.