We develop a common theoretical framework for combining classifiers which use distinct pattern representations and show that many existing schemes can be considered as special cases of compound classification where all the pattern representations are used jointly to make a decision. An experimental comparison of various classifier combination schemes demonstrates that the combination rule developed under the most restrictive assumptions-the sum rule-outperforms other classifier combinations schemes. A sensitivity analysis of the various schemes to estimation errors is carried out to show that this finding can be justified theoretically.

On combining classifiers

It has been demonstrated that combining the decisions of several classifiers can lead to better recognition results. The combination can be implemented using a variety of strategies, among which majority vote is by far the simplest, and yet it has been found to be just as effective as more complicated schemes in improving the recognition results. This paper examines the mode of operation of the majority vote method in order to gain a deeper understanding of how and why it works, so that a more solid basis can be provided for its future applications to different data and/or domains. In the course of our research, we have analyzed this method from its foundations and obtained many new and original results regarding its behavior. Particular attention has been directed toward the changes in the correct and error rates when classifiers are added, and conditions are derived under which their addition/elimination would be valid for the specific objectives of the application. At the same time, our theoretical findings are compared against experimental results, and these results do reflect the trends predicted by the theoretical considerations.

/pdf/application-of-majority-voting-to-pattern-recognition-an-40e00ps8pl.pdf

Application of majority voting to pattern recognition: an analysis of its behavior and performance

For pattern recognition, when a single classifier cannot provide a decision which is 100 percent correct, multiple classifiers should be able to achieve higher accuracy. This is because group decisions are generally better than any individual's. Based on this concept, a method called the "Behavior-Knowledge Space Method" was developed, which can aggregate the decisions obtained from individual classifiers and derive the best final decisions from the statistical point of view. Experiments on 46451 samples of unconstrained handwritten numerals have shown that this method achieves very promising performances and outperforms voting, Bayesian, and Dempster-Shafer approaches. >

A method of combining multiple experts for the recognition of unconstrained handwritten numerals

1. Spoken language input Ronald Cole, Victor Zue, Wayne Ward, Melvyn J. Hunt, Richard M. Stern, Renato De Mori, Fabio Brugnara, Salim Roukos, Sadaoki Furui and Patti Price 2. Written language input Joseph Mariani, Sargur N. Srihari, Rohini K. Srihari, Richard G. Casey, Abdel Belaid, Claudie Faure, Eric Lecolinet, Isabelle Guyo, Colin Warwick and Rejean Plamondon 3. Language analysis and understanding Annie Zaenen, Hans Uszkoreit, Fred Karlsson, Lauri Karttunen, Antonio Sanfilippo, Stephen F. Pulman, Fernando Pereira and Ted Briscoe 4. Language generation Hans Uszkoreit, Eduard Hovy, Gertjan van Noord, Gunter Neumann and John Bateman 5. Spoken output technologies Ronald Cole, Yoshinori Sagisaka, Christophe d'Alessandro, Jean-Sylvain Lienard, Richard Sproat, Kathleen R. McKeown and Johanna D. Moore 6. Discourse and dialogue Hans Uszkoreit, Barbara Grosz, Donia Scott, Hans Kamp, Phil Cohe and Egidio Giachin 7. Document processing Annie Zaenen, Per-Kristian Halvorsen, Donna Harman, Peter Schauble, Alan Smeaton, Paul Jacobs, Karen Sparck Jones, Robert Dale, Richard H. Wojcik and James E. Hoard 8. Multilinguality Annie Zaenen, Martin Kay, Christian Boitet, Christian Fluhr, Alexander Waibel, Yeshwant K. Muthusamy and A. Lawrence Spitz 9. Multimodality Joseph Mariani, James L. Flanagan, Gerard Ligozat, Wolfgang Wahlster, Yacine Bellik, Alan J. Goldschen, Christian Benoit, Dominic W. Massaro and Michael M. Cohen 10. Transmission and storage Victor Zue, Isabel Trancoso, Bishnu S. Atal, Nikil S. Jayant and Dirk Van Compernolle 11. Mathematical methods Ronald Cole, Hans Uszkoreit, Steve Levinson, John Makhoul, Aravind Joshi, Herve Bourlard, Nelson Morgan, Ronald M. Kaplan and John Bridle 12. Language resources Ronald Cole, Antonio Zampolli, Eva Ejerhed, Ken Church, Lori Lamel, Ralph Grishman, Nicoletta Calzolari, Christian Galinski and Gerhard Budin 13. Evaluation Joseph Mariani, Lynette Hirschman, Henry S. Thompson, Beth Sundheim, John Hutchins, Ezra Black, Margaret King, David S. Pallett, Adrian Fourcin, Louis C. W. Pols, Sharon Oviatt, Herman J. M. Steeneken and Junichi Kanai Glossary Citation index Index.

Survey of the State of the Art in Human Language Technology

The problem of classifier combination is considered in the context of the two main fusion scenarios: fusion of opinions based on identical and on distinct representations. We develop a theoretical framework for classifier combination for these two scenarios. For multiple experts using distinct representations we argue that many existing schemes such as the product rule, sum rule, min rule, max rule, majority voting, and weighted combination, can be considered as special cases of compound classification. We then consider the effect of classifier combination in the case of multiple experts using a shared representation where the aim of fusion is to obtain a better estimate of the appropriatea posteriori class probabilities. We also show that the two theoretical frameworks can be used for devising fusion strategies when the individual experts use features some of which are shared and the remaining ones distinct. We show that in both cases (distinct and shared representations), the expert fusion involves the computation of a linear or nonlinear function of thea posteriori class probabilities estimated by the individual experts. Classifier combination can therefore be viewed as a multistage classification process whereby thea posteriori class probabilities generated by the individual classifiers are considered as features for a second stage classification scheme. Most importantly, when the linear or nonlinear combination functions are obtained by training, the distinctions between the two scenarios fade away, and one can view classifier fusion in a unified way.

/pdf/combining-classifiers-a-theoretical-framework-4ruzg0fhdh.pdf

Combining classifiers: A theoretical framework

The authors present a conceptual framework for solving the task of document analysis, which, in essence, consists in the conversion of the document's pixel representation into an equivalent knowledge network representation holding the document's content and layout. Starting on the pixel level, the formation of elementary geometric objects on which layout analysis as well as the definition of character objects is based is described. Character recognition accomplishes the mapping from geometric object to character meaning in ASCII representation. On the next level of abstraction words are formed and verified by contextual processing. Modeled knowledge about complete documents and about how their constituents are related to the application forms the highest level of abstraction. The various problems arising at each stage are discussed. The dependencies between the different levels are exemplified and technical solutions put forward. >

Document analysis—from pixels to contents

Two different approaches are described to combine the results of different classifiers. The first approach is based on the Dempster/Shafer theory of evidence and the second one is a statistical approach with some assumptions on the input data. Both approaches were tested for user-dependent recognition of on-line handwritten characters. >

A comparison of two approaches for combining the votes of cooperating classifiers

A method of representing connected components in multivalued images is proposed, and an efficient, sequential, one-pass algorithm for generating the border-line chain code for each component is introduced. Simultaneously, the algorithm produces for each component a list of all neighboring components that are encountered during a walk around the border line, thus providing full topological information. The number of different pixel values does not affect performance. >

E. Mandler

Papers

Document analysis—from pixels to contents

A comparison of two approaches for combining the votes of cooperating classifiers

One-pass encoding of connected components in multivalued images

Document analysis-from pixels to contents