Reinvigorate Kernel-based algorithms with Isolation Kernel
Abstract
The need to manually design a suitable kernel for the task at hand has been the bottleneck ofkernel-based algorithms for about two decades. This project aims to create a new datadependent kernel which will be a breakthrough in easing the bottleneck. The isolationmechanism which produces the proposed Isolation Kernel will automatically adapt to theunderlying data distribution directly from a given dataset. It aims to simply replace the dataindependent kernel in Support Vector Machines (SVM), leaving the rest of the SVM procedureunchanged. It can be applied to wider applications and has significantly lower time and spacecomplexities than existing methods such as distance metric learning, multiple kernel learning,conformal transformation and Random Forest kernel. This is because these existing methodsrequire explicit learning (some require changes in the SVM procedure) and most need classinformation; but the proposed method needs neither. This project will verify the effectivenessof the proposed kernel using SVM in classification and clustering tasks.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 24, 2019
- Source ID
- FA23861814032
Entities
People
- Kai Ming Ting
Organizations
- Air Force Office of Scientific Research
- Federation University Australia
- United States Air Force