南航暑期国际课程大数据可视化第8讲2综述.ppt

Multidimensional collaborative lossless visualization: experimental study Vladimir Grishin 1 Boris Kovalerchuk 2 1 View Trends International 2 Dept. of Computer Science, Central Washington University Agenda Motivation and Approach: Collaborative visual shape pattern recognition to enhance machine learning Visual representation: Stars of Collocated Pair Coordinates (CPC) Applications of CPC Stars Experiments Conclusion Motivation and Approach Collaborative visual shape pattern recognition to enhance machine learning Lossy visualization, such as n-D reduction, principal component, n-D scaling, display small part of information contained in data and that small part is of unknown importance. 2. Popular lossless displays, such as parallel coordinates, and RadViz use poorly human capability of visual recognition of patterns [1,5,6]. 3. Many available visualizations are not scalable beyond small samples. Our Approach: Analytical and experimental exploration of structures of multidimensional data by means of new lossless n-D visualization methods providing effective collaborative usage of human capabilities in visual recognition of 2-D shapes [1,3,4] Visualization and Shape Perception Shape perception supplies over 90% of data for human pattern recognition. For visual pattern recognition humans can detect, compare, describe and combine into a multilevel hierarchy figures with hundreds of local features [1.3] such as: concave, convex, angle, wave, etc, and attributes of these features, size, orientation, location, etc. Leverage millions of years of visual system evolution vs. 70 years of computer technology evolution. Shape Vision Star–polar display of data vector X Each attribute xi of n-D vector X is mapped into polar ray length (“rose” of {xi}) after shifting the origin of each coordinate to provide non-negative values of all xi. Stars of two data vectors (points) For n in t