Mode-Based Clustering with Applications to Information Visualization

• Activity Number: 189
• Type: Invited
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jia Li*+ and Xiaolong Zhang
• Companies: Penn State and Penn State
• Keywords: mode association clustering ; Gaussian mixture ; visualization

Abstract

Gaussian mixtures used for clustering continuous data imply that each cluster has an elliptical shape, an assumption often violated in real-world data sets. A recently developed method addresses this issue by partitioning data according to whether they can be brought to the same mode via an ascending path, assuming the density function is in the form of a mixture. The correspondence between a mixture component and a cluster no longer holds under the new paradigm. To treat large-scale real-world data, we have developed a strategy to significantly speed up the mode based clustering method. Leveraging this state-of-the-art clustering method, we have developed an interactive visualization system to enhance the presentation of curves. Moreover, a summarization method is developed for meteorology cloud maps used in weather prediction.