×

Threshold optimization for weighted voting classifiers. (English) Zbl 1043.90018

Summary: Weighted voting classifiers considered in this paper consist of \(N\) units each providing individual classification decisions. The entire system output is based on tallying the weighted votes for each decision and choosing the one which has total support weight exceeding a certain threshold. Each individual unit may abstain from voting. The entire system may also abstain from voting if no decision support weight exceeds the threshold. Existing methods of evaluating the reliability of weighted voting systems can be applied to limited special cases of these systems and impose some restrictions on their parameters. In this paper a universal generating function method is suggested which allows the reliability of weighted voting classifiers to be exactly evaluated without imposing constraints on unit weights. Based on this method, the classifier reliability is determined as a function of a threshold factor, and a procedure is suggested for finding the threshold which minimizes the cost of damage caused by classifier failures (misclassification and abstention may have different price.) Dynamic and static threshold voting rules are considered and compared. A method of analyzing the influence of units’ availability on the entire classifier reliability is suggested, and illustrative examples are presented.

MSC:

90B25 Reliability, availability, maintenance, inspection in operations research
62H30 Classification and discrimination; cluster analysis (statistical aspects)
PDFBibTeX XMLCite
Full Text: DOI

References:

[1] and Combining a fuzzy classifier with classifiers based on statistic moments, Proc 6th UK Mechatronics Forum Int Conf, 1998, pp. 607-611.
[2] Avizienis, IEEE Trans Comput C-20 pp 1312– (1971)
[3] and N-version programming: A fault tolerance approach to reliability of software operation, Proc Int Symp Fault-Tolerant Comput, Toulouse, France, 1978, pp. 3-9.
[4] Deogun, J Am Soc Inf Sci 49 pp 423– (1998)
[5] and Voting in clustering and finding the number of clusters, Proc Int ICSC Symp Adv Intell Data Anal (AIDA 99), ICSC, Academic Press, New York, 1999, pp. 291-296.
[6] Fan, ASME J Eng Ind 117 pp 412– (1995)
[7] Ishii, Syst Comput Jpn 27 pp 25– (1996)
[8] Kumar, IEEE Trans Reliab 40 pp 593– (1991)
[9] and ?Combining neural network voting classifiers and error correcting output codes,? MEASUREMENT ’97, and (Editors), Institute of Measurement Science, Slovak Academy of Sciences, Bratislava, Slovak Republic, 1997, pp. 266-269.
[10] Levitin, IEEE Trans Reliab 49 pp 71– (2000)
[11] Levitin, IEEE Trans Reliab 47 pp 165– (1998)
[12] Levitin, Reliab Eng Syst Safety 65 pp 271– (1999)
[13] Levitin, Reliab Eng Syst Safety 64 pp 33– (1999)
[14] Levitin, Reliab Eng Syst Safety 71 pp 131– (2001)
[15] Nordmann, IEEE Trans Reliab 48 pp 42– (1999)
[16] Parhami, IEEE Trans Reliab 43 pp 617– (1994)
[17] Pham, Reliab Eng Syst Safety 63 pp 13– (1999)
[18] and Numerical recipes in C. The art of scientific computing, Cambridge University Press, New York, 1992. · Zbl 0845.65001
[19] Ushakov, Sov J Comput Syst Sci 24 pp 118– (1986)
[20] ?Probabilistic logics and the synthesis of reliable organisms from unreliable components,? Automata studies (Annals of Mathematics Studies, No. 34), and (Editors), Princeton University Press, Princeton, NJ, 1956, pp. 43-98.
[21] and A voting scheme for cluster algorithms, Proc 4th Int Workshop Neural Networks Appl, and (Editors), Otto-von-Guericke University of Magdeburg, Germany, 1999, pp. 31-37.
[22] Wu, IEEE Trans Reliab 43 pp 327– (1994)
This reference list is based on information provided by the publisher or from digital mathematics libraries. Its items are heuristically matched to zbMATH identifiers and may contain data conversion errors. In some cases that data have been complemented/enhanced by data from zbMATH Open. This attempts to reflect the references listed in the original paper as accurately as possible without claiming completeness or a perfect matching.