en
0.25
0.5
0.75
1.25
1.5
1.75
2
Metric Anomaly Detection Via Asymmetric Risk Minimization
Published on Oct 17, 20112974 Views
We propose what appears to be the first anomaly detection framework that learns from positive examples only and is sensitive to substantial differences in the presentation and penalization of normal v
Related categories
Chapter list
Anomaly Detection Via Asymmetric Risk Minimization00:00
Philosophy of Anomaly Detection00:24
Cost-Sensitive One-Class Anomaly Detection01:01
Problem Definition02:34
Common Modeling Assumption: Euclidean Space03:21
What About Metric Space?04:53
Background06:08
Metric Space06:15
Binary Classification for Metric Data06:52
Computational Efficiency08:23
Doubling Dimension08:52
Generalization Bound in Metric Space10:27
Model Assumptions11:31
Positive Points11:37
Negative Points11:49
Additional Positive Points12:00
Separation Distance12:16
Anomaly Detection via Asymmetric Risk Minimization12:32
Various Models of Uncertainty12:38
Instead of Generalization Error - Asymmetric Risk13:21
1st Case: A Known Separation Distance - 114:13
1st Case: A Known Separation Distance - 214:43
Bound the False Alarm Rate15:37
Bound the Risk16:47
2nd Case: We Have a Prior17:09
Choosing the Optimal Separation Distance17:21
3rd Case: No Explicit Prior17:53
Estimating the False-Alarm Component - 118:52
Estimating the False-Alarm Component - 219:50
Estimating the Missed Anomaly Components20:20
False Alarms Are Possible21:28
Missed Anomalies Are Possible21:31
Empirical Experiment21:33
Classification Results and The Incurred Classification Cost21:38
Thank You23:07