Model-fitting in the presence of outliers

We study the problem of parametric model-fitting in a finite alphabet setting. We characterize the weak convergence of the goodness-of-fit statistic with respect to an exponential family when the observations are drawn from some alternate distribution. We then study the effects of outliers on the model-fitting procedure by specializing our results to $\epsilon$-contaminated versions of distributions from the exponential family. We characterize the sensitivity of various distributions from the exponential family to outliers, and provide guidelines for choosing thresholds for a goodness-of-fit test that is robust to outliers in the data.


Published in:
2011 IEEE International Symposium On Information Theory Proceedings (ISIT), 1598-1602
Presented at:
IEEE International Symposium on Information Theory (ISIT 2011), Saint-Petersburg, Russia, July, 31 - August, 5, 2011
Year:
2011
Publisher:
IEEE Service Center, 445 Hoes Lane, Po Box 1331, Piscataway, Nj 08855-1331 Usa
Keywords:
Laboratories:




 Record created 2011-05-17, last modified 2018-03-17

n/a:
Download fulltext
PDF

Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)