Intelligence softwareStrengthening national security by improving intelligence software

An intelligence analyst hunting for answers in a sea of data faces steep challenges: She must choose the right search terms, identify useful results, and organize them in a way that reveals new connections. 

Making that process quicker and more intuitive could yield faster answers to key national security questions, which is why a research group at Virginia Tech is collaborating with Fairfax-based defense company General Dynamics Mission Systems on intelligence software that allows analysts to interact more closely with their data.

According to Chris North, a professor of computer science in the College of Engineering and the associate director of the Discovery Analytics Center, analysts currently have to approach huge data sets with independent, consecutive searches.

“You search, and then you read. And you read and you read and you read. And then you might figure out from all that reading something else you might search for, and then you do that, and it’s a slow, painful iterative process,” North said.

Virginia Tech says that North’s research group is developing software that uses a visual interface and computer learning algorithms to allow to the analyst’s interactions with the data to guide future searches.

Demonstrating the system, computer science doctoral student Michelle Dowling, from Grand Rapids, Michigan, enters the name of a person of interest in the search field, and a constellation of nodes pops up on the screen. Each node represents a document containing the name Dowling searched for; the documents belong to a data set the government uses to train intelligence analysts.

The size of each node represents the algorithm’s assessment of that document’s relevance, and the distance between any two nodes reflects the similarity of those two results to each other.

Because of the system’s graphical interface, the analyst can move two nodes closer together to indicate any important similarities; the results will rearrange themselves to show which documents are most relevant based on that analysis.

For example, bringing together nodes for a restaurant receipt and a plane ticket may suggest that a particular trip might be important; the algorithm can pull in other results related to that date or location.

And past search terms will be weighted more heavily when they pop up in the results of future searches.