|
Activity Number:
|
478
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2007 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #309880 |
|
Title:
|
Unsupervised Quantile Learning To Identify Principal Direction of Anomaly
|
|
Author(s):
|
Aijun Zhang*+ and Agus Sudjianto and Ming Yuan
|
|
Companies:
|
University of Michigan and Bank of America and Georgia Institute of Technology
|
|
Address:
|
439 West Hall, Ann Arbor, MI, 48109,
|
|
Keywords:
|
Unsupervised learning ; Quantile ; Outlier detection
|
|
Abstract:
|
We propose a quantile learning framework to tackle some unsupervised tasks. It is formulated as a maximin problem with double convex objectives. This paper studies the extreme quantile case, in order to identify the principal direction of anomaly and detect outliers. It is shown that our outlier peeling algorithm can be viewed as a sequential variate of the classical Mahalanobis depth approach.
|
