In the days of Sherlock Holmes, solving a crime depended entirely on a person's wits. These days, with fictional cops like those of "CSI" dominating popular culture, it can seem like solving a crime depends entirely on computers.

But Byron Marshall thinks a combination of human intuition and artificial intelligence (AI) could offer a streamlined way to help investigators get to the bottom of things.

Marshall, an assistant professor in information management at Oregon State University's School of Business, has been working on a method that allows investigators to identify people of interest in a case based on relationships.

Finding such networks can help lead police to a person involved with a particular crime, even if the individual's criminal record doesn't indicate involvement.

Marshall's research will be published in a future issue of the Journal of the American Society for Information Science and Technology.

Marshall said the technique was motivated by the idea that he could make things easier for investigators.

In Arizona, where he got his Ph.D., he watched law enforcement agents use a database called Cop Link. He saw room for improvement.

"They had access to lots of data, but at the end of the day the detectives were reduced to doing queries by what I thought were very simplistic methods," Marshall said.

Working with Siddharth Kaza and Hsinchun Chen of the University of Arizona's Artificial Intelligence lab, Marshall helped come up with an algorithm to make the process more efficient.

Finding associations

The method uses both human expertise and AI to identify known associations between people.

A crime analyst can look at thousands of files before coming up with a few people to talk to in connection with a case, Marshall said. With the algorithm, the analyst still finds those people - but a lot faster.

Standard AI would find a pattern and identify a suspect who fits that pattern.

"We turn that on its head," Marshall said.

Instead, the process, which uses only police data, starts with a known suspect involved in a crime, he said, then moves to known associates.

Rather than trying to mirror human knowledge, Marshall said, the method helps take advantage of it.

"A detective can say 'I'm more interested in people if they have done this and this and this,'" he said. "They can customize that to a particular case."

Experts see patterns other would miss, Marshall said, and know when those patterns matter.

"That sort of thinking is difficult to include in a computational program."

The interactive nature also means information can be added as it is discovered - say, if a detective finds out a suspect had a previously unknown sibling.

And unlike AI, the computer is not expected to do all the work.

"We just try to help the computer do what it does well and let the person do what they do well," Marshall said.

For one thing, Marshall's method doesn't just find relationships, it evaluates them.

"The approach recognizes that different kinds of relationships are more or less important," he said.

"Two people in a domestic violence case probably know each other pretty well," Marshall said. "Two people who were both interviewed as witnesses in a traffic accident probably don't."

But he emphasizes that it's not about making people guilty by association.

"We're really concerned about that," he said.

The information, which comes solely from police data, may help investigators identify a suspect, but it doesn't produce an arrest warrant in and of itself.

"They're going to find all these people anyway; we're just helping them save some time," Marshall said.

"That's much different than saying 'Look through and find anybody who's behind on their credit card payments,'" he said. "It does not scan through a data bank looking for a pattern or profile."

And, the method could be widely employed without disclosing private information.

"The analysis can operate on relatively nonintrusive data," he said.

The goal, he said, is to share enough data to be useful to an investigation, without compromising privacy.

The approach is new, but the research started seven years ago. Another decade or so before that, the researchers began working with law enforcement, particularly Tucson Police Department Crime Analyst Kathy Martinjak, to record the necessary data.

But the effort is paying off. Though there is not yet a working prototype to employ the technique, the research has potential to be applied to fields other than law enforcement.

Bio-medical research, Marshall said, might benefit tremendously.

For instance, one gene might not be identified with a particular disease, but the algorithm could reveal if the gene is connected to other genes that are.

The technique also could be used to help teachers select curriculum to teach to specific standards.

Marshall is excited by the possibilities.

"I think the really cool thing is leveraging someone's intuition," he said.