Voulgaris, Z. and Mirkin, Boris (2010) Choosing a discernibility measure for reject-option of individual and multiple classifiers. International Journal of General Systems 39 (8), pp. 855-871. ISSN 0308-1079.Full text not available from this repository.
A novel method for evaluating the reliability of a classifier on a pattern is proposed based on the discernibility of a pattern's class against other classes from the pattern. Three measures of discernibility are proposed and experimentally compared with each other and with more conventional techniques based on the classification scores for class labels. The classification accuracy can be significantly enhanced through discernibility measures using the most reliable - 'elite' - patterns. It can be further boosted by forming an amalgamation of the elites of different classifiers. Improved performance is achieved at the price of rejecting many patterns. There are situations in which this price is worth paying - when the non-reliable predictions, however good, lead to the need for the manual testing of very cumbersome and complex technical devices or in diagnostics of human terminal diseases. Contrary to conventional techniques for estimating reliability, the proposed measures are applicable to small datasets as well as to datasets with complex class structures on which conventional classifiers show low accuracy rates.
|Keyword(s) / Subject(s):||pattern recognition, rejection, reliability elite, discernibility, combining classifiers|
|School or Research Centre:||Birkbeck Schools and Research Centres > School of Business, Economics & Informatics > Computer Science and Information Systems|
|Date Deposited:||26 May 2011 11:28|
|Last Modified:||17 Apr 2013 12:20|
Archive Staff Only (login required)