Cooperative Classifiers Rozita Dara Supervisor: Prof. Kamel Pattern Analysis and Machine Intelligence Lab University of Waterloo
Combining Classifiers Goals: Improve performance over constituent classifiers. Maximize information use. Obtain a reliable system. Challenges: Intelligent combination that exploits complementary information.
Problem What type of cooperation between classifiers is the most effective? What important criteria should be considered when designing a multiple classifier system? What combination method is the best for a specific problem?
Objectives Enhance understanding of the combination methods and their applications. Obtain insights into designing and developing new architectures. Examine the usefulness and efficiency of our finding for document categorization.
Proposed Approach A thorough understanding of cooperation among Multiple classifiers System components provides guidelines for optimization of the system. Different levels of sharing Training Level Feature Level Architecture Level Decision Level
Proposed Approach (cont ’ d) Training Level Sharing training patterns Sharing training algorithm Feature Level Sharing features Architecture Level Sharing information Decision Level Sharing classifiers ’ decision
Key Accomplishments Training Level Training Data: Disjoint, Overlapped, and identical Training Data Size small, medium, and large Data dimensionality small and large Type of data large interclass distances and small interclass distances Architectures ensemble and modular
Research in Progress Sharing training algorithm architectures Sharing at feature level overlapped, identical, disjoint Sharing at architecture level share information Sharing at decision level classifiers ’ output
Research in Progress (cont ’ d) The advantages of using multiple classifiers in document analysis have been realized in recent years. Document data high dimensional large number of classes large number of inputs patterns