Seminar - Kernelized Fuzzy Rough Sets: Characterizing Inconsistency in Classification
Thu., Nov. 10, 2016 2:00 p.m.
Location: ED 191
Speaker: Dr. Qinghua Hu - School of Computer Science and Technology, Tianjin University, Tianjin, China.
Title: Kernelized Fuzzy Rough Sets: Characterizing Inconsistency in Classification
Time & Date: Thursday, November 10, 2016 at 2:00pm
Place: ED 191
Abstract:Rough sets are now widely used to characterize the inconsistency of classification and regression tasks. Given a universe U, we can granulate and organize the elements in U with Relation R, and then apply the derived information granules to approximate any other granules defined in advance. In fact, most rough set models are named with the relation R, such as neighborhood rough sets, dominance rough sets and fuzzy preference rough sets. R determines the structure of the approximation space, and then has great impact on the computation of lower and upper approximations. However, there are few works systematically discussing the issue of relation generation. In this talk, we will build a bridge between kernel machines and rough sets, and show that a collection of kernel functions can be used to calculate the relations of objects. Especially, fuzzy equivalence relations can be generated with some symmetric and reflexive kernels. We integrate the kernel functions with rough sets and construct kernelized rough set models. Moreover, we design multi-kernel fuzzy rough sets to analyze multi-modality data, such as the mixture of audios, texts, images and videos. In addition, it is reported that fuzzy rough sets are sensitive to noisy samples. We develop some robust kernelized fuzzy rough set models to combat this challenge.
After the model of kernelized fuzzy rough sets is developed, we have a question what the model can be used to address. First, the model is used to find the boundary samples in classification learning. As we know, the classification surface is generally determined by these samples. It is very useful to find these boundary samples before training. Second, we define some statistical factors for evaluating features based on the proposed model, and we call it fuzzy dependency functions. Some efficient feature selection algorithms are designed based on these functions. Finally, we propose an interesting technique for fuzzy multi-label classification. All these proposed models and algorithms are tested with real-world tasks.
Qinghua Hu received the B.S., M.S., and Ph.D. degrees from Harbin Institute of Technology, Harbin, China. He then became a Post-Doctoral Fellow with the Department of Computing, Hong Kong Polytechnic University, Hong Kong. He joined Tianjin University in 2012 and is currently a Full Professor and the Vice Dean of the School of Computer Science and Technology, Tianjin University. He is also director of the Lab of Machine Learning and Data Mining. He has published over 100 journal and conference papers in the areas of granular computing-based machine learning, reasoning with uncertainty, pattern recognition, including IJCAI, AAAI, ICCV, IEEE Trans. on Fuzzy System, and so on.
Prof. Hu was acted as the Program Committee Co-Chair of the International Conference on Rough Sets and Current Trends in Computing in 2010, the Chinese Rough Set and Soft Computing Society in 2012 and 2014, and the International Conference on Rough Sets and Knowledge Technology and the International Conference on Machine Learning and Cybernetics in 2014,general Co-Chair of IJCRS 2015. Now he is organizing a special issue in Information Sciences entitled Granular Computing Based Machine Learning in the Era of Big Data. He will organize the China Conference of Machine Learning (CCML 2017) in Tianjin and act as a PC Chair.