« Seminar - Pattern Clustering and Data Grouping | Main | Seminar - "China Internet Market - 2006 Performance and 2007 Preview" »
January 15, 2007
Seminar - Pattern Clustering and Data Grouping
Title : Pattern Clustering and Data Grouping
Speaker : Andrew K.C. Wong
Pattern Intelligence Inc. and
Systems Design Engineering, University of Waterloo
Date : 11 December 2006 (Monday)
Time : 2:30 p.m. ¡V 3:30 p.m.
Venue : PQ703, Department of Computing,
The Hong Kong Polytechnic University
Abstract:
A basic task of machine learning and data mining is to automatically uncover patterns that reflect regularities in a data set. Today, data contents become more complex and diverse. They usually contain both nominal and ordinal data. Interesting information and relevant patterns might be scattered, entangled and spanning in various data subspaces. In the past, we have developed an algorithm known as Pattern Discovery to discover statistically significant patterns
effectively wherever they are in the database. Recently, we have developed a new method known as Pattern Clustering and Data Grouping (PCDG) which is able to cluster similar patterns into pattern clusters while grouping pattern-induced data into data groups. In that sense data associated with similar statistical patterns or rules are automatically organized into groups. Then the probabilistic characteristics of patterns in each data group and the relationship among them can be revealed.
PCDG takes a data set and the association patterns discovered by either pattern discovery or association rule mining as inputs. It induces a data group located in the data set for each association pattern. Using a similarity measure between data groups, it applies a clustering algorithm to simultaneously cluster similar association patterns and group the respective pattern-induced data into data groups. Data groups induced usually span different subspaces and could overlap each other if they contain common parts of the discovered patterns. Clustering terminates based on a stopping criterion. After pattern clustering and data grouping, the probabilistic characteristics of each data groups can be explicitly displayed. PGDG can further discover and represents the structural relations among data groups in the form of attributed hypergraphs, revealing the similarity and differences among data groups. Hence, knowledge trapped in the data can be unveiled and organized for interpretation and understanding. In the talk, supporting experiments on synthetic and real-world data, including automatic discovery of gene expression subgroups from microarray data of cancerous tissues and DNA splicing classes in DNA sequences will be reported.
Posted by khtsoi at January 15, 2007 12:26 PM