Systems Seminar
Transductive Anomaly Detection
Prof. Clay Scott
EECS Dept.
University of Michigan
Abstract
In machine learning, one formulation of the anomaly detection problem
is to build a detector based on a training sample consisting only on
nominal data. The standard approach to this problem has been to
declare anomalies where the density of the nominal data is low. This
approach is inductive in the sense that the detector is constructed
before any test data are observed. In this talk, we consider the
transductive setting where the unlabeled and possibly contaminated
test sample is also available at learning time. We argue that anomaly
detection in this transductive setting is naturally solved by a
reduction to a binary classification problem. In particular, an
anomaly detector with a desired false alarm rate can be achieved
through a reduction to Neyman-Pearson classification. Unlike the
inductive approach, this transductive approach yields detectors that
are optimal (e.g., statistically consistent) regardless of the
distribution on anomalies. Therefore, in anomaly detection, unlabeled
data have a substantial impact on the theoretical properties of the
decision rule. Time permitting, I will discuss other problems that
arise in transductive anomaly detection, including multiple testing
issues and estimating the proportion of anomalies in the test sample.
This is joint work with Gilles Blanchard.
Time and Place: Wed., Apr. 16, at 3:30 pm in 4610.
SYSTEMS SEMINAR WEB PAGE:
http://homepages.cae.wisc.edu/~gubner/seminar/schedule.html