Nuno Homem,

Mainroad

Abstract:

This work proposes algorithms and methods for individual behavior detection within very large populations. One will consider domains where individual behavior presents some stable characteristics over time, and where the individual actions can be observed through events in a data stream. Event patterns will be characterized and used as a proxy to individual behavior and actions. As in many domains, behavior does not remain static but evolves over time; one will therefore consider the sliding window model, making the assumption that behavior is stable during the considered time window.

This work will cover the detection of the specific characteristics of the individual and what distinguishes his behavior from that of all other individuals. Algorithms must have minimal memory footprint and scalability to cope with huge number of individuals. Providing and keeping results up to date in near real time is also a goal, as information is only useful for limited periods in many situations. Fortunately, approximate answers are usually adequate for most problems.

Some fast and compact methods for diversity analysis will be introduced both for unlimited time and for the sliding window model. Innovative algorithms will be proposed to describe and characterize the individual event patterns. Those algorithms will then be used to create an individual event fingerprint. Using that fingerprint one will be able to identify the individual even when the identification information is not available. Distinct uses of the fuzzy fingerprint concept will be presented for individual identification that might also be extended to specific behavior identification, classification, profiling, etc., with examples in several domains such as internet traffic analysis, telecommunications fraud detection and text authorship analysis.

 

Date: 2011-Oct-28     Time: 15:00:00     Room: 020


For more information: