Learning from Data Streams Using Kernel Adaptive Filtering

Abstract

A learning task is sequential if its data samples become available over time. Kernel adaptive filters (KAF) are sequential learning algorithms. There are two main challenges in KAF: (1) the lack of an effective method to determine the kernel-sizes in the online learning context; (2) how to tune the step-size parameter. We propose a framework for online prediction using KAF which does not require a predefined set of kernel-sizes; rather, the kernel-sizes are both created and updated in an online sequential way. Further, to improve convergence time, we propose an online technique to optimize the step-size parameter. The framework is tested on two real-world data sets, i.e., internet traffic and foreign exchange market. Results show that, without any specific hyperparameter tuning, our proposal converges faster to relatively low values of mean squared error and achieves better accuracy.

Keywords: Learning From Data Streams, Sequence Prediction, Kernel Adaptive Filters, Kernel Least Mean Square