Improving classification performance of support vector machines via guided custom kernel search

Abstract

Support Vector Machines (SVMs) deliver state-of-the-art performance in real-world applications and are established as one of the standard tools for machine learning and data mining. A key problem of these methods is how to choose an optimal kernel function. The real-world applications have also emphasized the need to adapt the kernel to the characteristics of heterogeneous data in order to boost the classification accuracy. Therefore, our goal is to automatically search a task specific kernel function. We use reinforcement learning based search mechanisms to discover custom kernel functions and verify the effectiveness of our approach by conducting an empirical evaluation with the discovered kernel function on MNIST classification. Our experiments show that the discovered kernel function shows significantly better classification performance than well-known classic kernels. Our solution will be very effective for resource constrained systems with low memory footprint which rely on traditional machine learning algorithms like SVMs for classification tasks.