In *Proceedings of Odyssey 2004: The Speaker and Language Recognition Workshop*, 2004.

Paper abstract bibtex

Paper abstract bibtex

Assessing whether two models are \em statistically significantly different from each other is a very important step in research, although it has unfortunately not received enough attention in the field of person authentication. Several performance measures are often used to compare models, such as \em half total error rates (HTERs) and \em equal error rates (EERs), but most being aggregates of two measures (such as the \em false acceptance rate and the \em false rejection rate), simple statistical tests cannot be used as is. We show in this paper how to adapt one of these tests in order to compute a confidence interval around one HTER measure or to assess the statistical significantness of the difference between two HTER measures. We also compare our technique with other solutions that are sometimes used in the literature and show why they yield often too optimistic results (resulting in false statements about statistical significantness).

@inproceedings{bengio:2004:odyssey:stats, author = {S. Bengio and J. Mari�thoz}, title = {A Statistical Significance Test for Person Authentication}, booktitle = {Proceedings of Odyssey 2004: The Speaker and Language Recognition Workshop}, year = 2004, url = {publications/ps/bengio_2004_odyssey_stats.ps.gz}, pdf = {publications/pdf/bengio_2004_odyssey_stats.pdf}, djvu = {publications/djvu/bengio_2004_odyssey_stats.djvu}, original = {2004/stat_test_odyssey}, idiap = {publications/pdf/rr03-83.pdf}, topics = {biometric_authentication}, web = {http://www.isca-speech.org/archive/odyssey_04/ody4_237.html}, abstract = {Assessing whether two models are {\em statistically significantly} different from each other is a very important step in research, although it has unfortunately not received enough attention in the field of person authentication. Several performance measures are often used to compare models, such as {\em half total error rates} (HTERs) and {\em equal error rates} (EERs), but most being aggregates of two measures (such as the {\em false acceptance rate} and the {\em false rejection rate}), simple statistical tests cannot be used as is. We show in this paper how to adapt one of these tests in order to compute a confidence interval around one HTER measure or to assess the statistical significantness of the difference between two HTER measures. We also compare our technique with other solutions that are sometimes used in the literature and show why they yield often too optimistic results (resulting in false statements about statistical significantness).}, categorie = {C}, }

Downloads: 0