2022-7: Distributional Kernels — an under-utilized resource

Date: Friday 5 Aug

Time: 9.00am – 10.00am

Speaker: Prof. Kai Ming Ting

Abstract:This seminar presents recent works on distributional kernels based on kernel mean embedding (KME). KME has a strong theoretical underpinning, and guarantees that the resultant kernel mean map is injective, i.e., the kernel mean maps of two distributions have their difference equals to zero if and only if the distributions are the same. Yet, KME’s applications have been less successfully so far. One key breakthrough is the identification of the root cause of KME’s (seemingly) failures, i.e., the use of Gaussian kernel. The seminar presents works, following this identification, that release the power of this under-utilized resource. The works demonstrate that the distributional kernels can solve long-standing problems, some of which have evaded decades of effort, in terms of efficiency and task-specific accuracy issues. These include point and group anomaly detections, clustering, and anomaly detections in trajectories, periodic time series and graphs/networks.

Bio: After receiving his PhD from the University of Sydney, Australia, Kai Ming Ting worked at the University of Waikato (NZ), Deakin University, Monash University and Federation University in Australia. He joined Nanjing University in 2020. Research grants received include those from National Science Foundation of China, US Air Force of Scientific Research (AFOSR/AOARD), Australian Research Council, Toyota InfoTechnology Center and Australian Institute of Sport. Professor Ting is one of the inventors of Isolation Forest, Isolation Kernel and Isolation Distributional Kernel. Isolation Forest is widely used in industries and academia. Isolation Kernel is a unique similarity measure which is derived from a dataset based on the same/similar isolation mechanism as Isolation Forest, and has no closed-form expression. Isolation Kernel and Isolation Distributional Kernel are the X-factor that enables many problems to be solved more effectively and efficiently than existing algorithms which rely on Gaussian kernel or Euclidean distance.