Two years ago, Nancy Kaup was a 31-year-old single mother who was frustrated with dating. She had spent six months on the website eHarmony, filled out a 400-question survey about herself, and begun receiving daily “matches”—profiles of men whom the site deemed compatible. But none of them worked out. She decided not to renew her subscription. Two days before her profile expired, however, a man named Jon Anthony signed up for the service.
Nancy showed up in Jon’s first round of suggested matches, and he contacted her. “He was my last match and I was his first,” she says. Their first date was at a wine tasting in Albuquerque, New Mexico, where they both live. Although it lasted only an hour or two, the next day Nancy told her friends at work that she had met her future husband. “I knew right away,” she says. “It’s weird, because I’m not usually like that.”
The online dating industry is bigger than ever. An estimated 40 million Americans are members of dating services offered over the Web or on mobile devices, and in China the number has grown to 140 million people. But matching up millions of members is a major technological challenge as well as an emotional one. While some sites simply let users browse for dates, many now offer some kind of system, if only to make suggestions. And companies in this competitive market are in hot pursuit of ways to make those suggestions more sophisticated and personalized. To do that, they are deploying machine-learning algorithms that are adapted from completely different types of online shopping.
Joseph Essas, vice president of technology for eHarmony, was lured to the company from Yahoo three years ago. Since then, he has developed and implemented a new layer of predictive matching algorithms that are based on Yahoo’s system for targeting advertising to specific users who have revealed preferences and behaviors over time. The matchmaking software gathers 600 data points for each user, including how often they log in, who they search for, and what characteristics are shared by the people they actually contact.
According to Essas, eHarmony has used this information to predict how likely it is that two people will engage in conversation, which helps determine which matches will be suggested on any given day. “How do we get people talking to each other to recognize their commonalities?” he asks. The new software, he says, gets more such conversations started, “with 34 percent more back-and-forth communication compared to a year ago.”
While most of these new techniques were installed after Nancy first met Jon, eHarmony has built stories like theirs into their model, as these are the kind of matchups the company aims for. Jon and Nancy were engaged within two months, and in five more months they were married. Now they have a baby on the way.
Adaptive algorithms are a powerful tool in online dating because what people say they want and how they actually behave are different things. Some people say they’re looking for a nonsmoker, for instance, but in practice they’ll date a smoker who fits their other criteria. Basing recommendations on behavior also translates into fewer time-consuming questions. “We can piece things together without having to ask you,” says Sam Yagan, CEO of OKCupid, a free online dating site. Often, the process can tease out information that would be impossible to get through a questionnaire. OKCupid, for instance, uses communication and ratings from other users to assign an attractiveness value to each member. When you are shown matches, says Yagan, they tend to fall within a range of attractiveness that matches your own.
These kinds of approaches are different from what was used before. For more than a decade, for example, eHarmony has beenusing an extensive questionnaire to characterize each member according to 29 “dimensions” of personality, identified by research on married couples as being important for long-term compatibility. Weighing which traits work well together and which do not, it offers members daily matches within certain user-selected criteria, like age, location, and religious beliefs.
But the new techniques are based not on questionnaires but on other kinds of “recommendation engines,” like those used by Netflix and Amazon, says Gavin Potter, chief technology officer of IntroAnalytics, a company that develops software for both e-commerce and dating sites. In the future, it could work the other way, too: matchmaking algorithms may help improve other kinds of online commerce. While shopping for a book and shopping for love do have some things in common, says Potter, one difference is that dating recommendations are bidirectional. “The object you’re recommending has got to be interested as well,” he says. If every person were shown the 10 hottest people on the site, the system wouldn’t work.
For all these companies, one major hurdle stands in the way of improving the algorithms: measuring success. It’s hard to know whether members find love after they take their interaction off the site.
Plenty of Fish, one of the largest dating websites in the United States, has taken the extra step of asking members who leave the site whether they entered a relationship with another member, says the company’s CEO, Markus Frind. This information is added to the company’s predictive model, which also includes information from personality tests and user behavior.
To figure out rates of success, OKCupid is currently analyzing online messages for 10-digit strings of numbers, figuring that exchanging phone numbers is a sign of success. Meanwhile, eHarmony is conducting a longitudinal study to follow a cohort of couples through five years of marriage, to see if those matched on eHarmony are truly more compatible. But unfortunately for singles ultimately hoping to find a soul mate online the way Nancy and Jon Anthony did, it’s ultimately impossible to know whether it’s any algorithm that’s doing the trick—or whether it’s a mix of old-fashioned instinct and good luck.