Kang Zhao, assistant professor of management sciences in the Tippie College of Business at the University of Iowa, and doctoral student Xi Wang were part of a team that developed an algorithm for dating sites that uses a person's contact history to recommend partners with whom they may be more amorously compatible.
It's similar to the model Netflix uses to recommend movies users might like by tracking their viewing history, the researchers said.
The team used data provided by a popular commercial online dating company. They looked at 475,000 initial contacts involving 47,000 users in two U.S. cities over a 196-day span. Of the users, 28,000 were men and 19,000 were women, and men made 80 percent of the initial contacts, Zhao said.
The data suggested only about 25 percent of those initial contacts were actually reciprocated. To improve this rate, Zhao's team developed a model that combines two factors to recommend contacts: a client's tastes -- determined by the types of people the client has contacted -- and attractiveness/unattractiveness -- determined by how many of those contacts are returned and how many are not.
The combinations of taste and attractiveness do a better job of predicting successful connections than relying on information that clients enter into their profiles, which can be misleading, or clients might not know themselves well enough to know their own tastes in the opposite sex, Zhao theorized.
A man who says on his profile he likes tall women might be approaching mostly short women, even though the dating website will continue to recommend tall women.
Zhao said the existing model leads to a return rate of about 25 percent, but the team's model could improve such returns by 44 percent.
The researchers said they found their model performed the best for males with athletic body types connecting with females with athletic or fit body types, and for females who indicate they "want many kids." The model also worked best for users who upload more photos of themselves.
The findings are scheduled to the published in IEEE Intelligent Systems and are available online at arxiv.org/pdf/1311.2526v1.pdf.