| Date | Speaker | Topic | |
|---|---|---|---|
| March 5th, 3:30pm | Diane Lambert, Bell Labs | Dynamic thresholding for network monitoring | |
| April 9th, 2:30pm | David Holmes, College of New Jersey | TBA | |
| May 7th, 11am | Tony Jebara, Columbia University | TBA | |
| May 7th, 3:30pm | Mark Goldberg, RPI | TBA |
Streams of counts on traffic volume and processing errors are continually compared to thresholds to detect network degradation. Thresholds are usually set by hand, which is tedious and error-prone, or by relying on normality assumptions, which gives unacceptably large false alarm rates. This talk will describe a simple, statistically principled approach to automated thresholding that starts from the statistical properties of the counts and then considers algorithms that meet the computing constraints. Longterm trends, cyclical patterns, outliers, and missing values are automatically accommodated in the thresholds. Just as importantly, events (periods of anomalous behavior) can be prioritized in terms of severity and duration, even though count distributions differ over time. Performance on real and simulated data will be described. (Joint work with Chuanhai Liu.)