Scientific Computing Seminar

Date:
Monday, April 25, 2005
Time:
1:00pm-2:00pm
Location:
50F-1647
Seminar Speaker:
Tsuyoshi Ide
IBM Research, Tokyo Research Lab.
http://www.research.ibm.com/trl/people/ide/index_e.htm
Title:
A spectral approach to anomaly detection in computer systems
Abstract:
Recently, knowledge discovery from structural data has attracted much attention as a new area of data mining. As an extension of the new area, I will talk about an approach to the issue of online anomaly detection for Web-based computer systems, which can be modeled as a graph with time-dependent weights.

Our method is divided into two parts: First, we extract effective information from the graph by using the principal eigenvector of the eigenclusters of the graph, so that the problem is reduced to that of outlier detection from a sequence of directional data.

Second, we introduce an online algorithm to calculate the threshold value for an anomaly measure based on the von Mises-Fisher distribution. Examples on a web server will be presented.

Sponsor of Seminar:
Chris Ding
Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov