|
|
|
Subdueing RHSEG: The Marriage of Graph Based Knowledge Discovery (Subdue) with Image Segmentation Hierarchies (from RHSEG) for Data Analysis, Data Mining and Knowledge Discovery
|
|
|
| It is proposed to design and implement the integration of the Subdue graph based knowledge discovery system, developed at the
University of Texas Arlington and Washington State University, with image segmentation hierarchies produced by the RHSEG
software, developed at NASA GSFC, and perform pilot demonstration studies of data analysis, mining and knowledge discovery on
NASA data. Subdue represents a method for discovering substructures in structural databases. Subdue is devised for general-purpose
automated discovery, concept learning, and hierarchical clustering, with or without domain knowledge. Hence, the method can be
applied to many structural domains. As an unsupervised algorithm, Subdue searches for a substructure, or subgraph of the input
graph, that best compresses the input graph. As a supervised learning algorithm, Subdue can be provided with known instances of
target concepts and can use this information to learn a graph concept. The graph concept can be used on new data to determine its
class value. Substructure discovery using Subdue has yielded expert-evaluated significant results in domains including terrorist
activity, predictive toxicology, network intrusion detection, earthquake analysis, web structure mining, and protein data analysis.
Subdue was developed by Co-I, Diane J. Cook, and her colleague, Lawrence B. Holder. For Subdue to be effective in finding patterns
in imagery data, the data must be abstracted up from the pixel domain. An appropriate abstraction of imagery data is a segmentation
hierarchy: a set of several segmentations of the same image at different levels of detail in which the segmentations at coarser levels of
detail can be produced from simple merges of regions at finer levels of detail. The RHSEG program, a recursive approximation to a
Hierarchical Segmentation approach (HSEG), can produce segmentation hierarchies quickly and effectively for a wide variety of
images. Both RHSEG and HSEG were developed at NASA GSFC by the PI, James C. Tilton. The HSEG/ RHSEG approaches are
based on region growing, and include a provision for merging spatially disjoint regions into aggregate regions. This region aggregation
leads to a compact abstraction of even very complex images with very faithful representation of spatial detail. This proposed work is an
innovative combination of two established information technologies which will significantly improve our ability to extract and/or
discover information from imagery data. The proposed work also includes a number of demonstration projects designed to illustrate
the significant potential this combination of technologies has to increase the productivity of NASA's Science Mission Directorate
research endeavors, fostering collaborations across a wide range of space, Earth and computer science disciplines. |
|
|
