Self-adaptive kernel machine: online clustering in RKHS

This paper presents a new online clustering algorithm (called SAKM) that is developed to learn continuously evolving clusters from non-stationary data. The SAKM algorithm is based on SVM methods with kernel trick in reproducing Hilbert space, and uses a fast incremental learning procedure to take into account model changes over time. Dedicated to online clustering in multi-class environment, the algorithm is based on an unsupervised learning process with self-adaptive abilities. The SAKM learning process is based on a specific kernel-induced similarity measure and is designed in four main stages: Creation with an initialisation procedure, adaptation, fusion and elimination. In addition to its new properties, the SAKM algorithm is attractive to be very computationally efficient and to provide good performances in online applications. After a comparison with NORMA and Gentile' ALMA algorithms, some experiments are presented to illustrate the capacities of our algorithm for online clustering of non-stationary data in multi-class environment.