Incremental kernel spectral clustering for online learning of non-stationary data

In this work a new model for online clustering named Incremental kernel spectral clustering (IKSC) is presented. It is based on kernel spectral clustering (KSC), a model designed in the Least Squares Support Vector Machines (LS-SVMs) framework, with primal-dual setting. The IKSC model is developed to quickly adapt itself to a changing environment, in order to learn evolving clusters with high accuracy. In contrast with other existing incremental spectral clustering approaches, the eigen-updating is performed in a model-based manner, by exploiting one of the Karush–Kuhn–Tucker (KKT) optimality conditions of the KSC problem. We test the capacities of IKSC with some experiments conducted on computer-generated data and a real-world data-set of PM10 concentrations registered during a pollution episode occurred in Northern Europe in January 2010. We observe that our model is able to precisely recognize the dynamics of shifting patterns in a non-stationary context.