How does the matching algorithm of the popular dating service suggest potential mates?
“Find Love. Guaranteed.” –Match.com tagline
I have a friend–let’s call her “Ruby”–whose dating life has lately experienced a dry spell. Worse than a dry spell, actually–more like a dry spell interrupted by intermittent acid rain. Things reached a crisis late one night, and in a fit of defeated desperation, she got out her credit card and pressed the button that sent $39.99 to Match.com, securing her a one-month membership with the online dating service. She told me of her decision the next day, and began a month-long quest for love through online dating. Naturally, I was suspicious of the whole endeavor from the start, but at the time I could not adequately explain why Match.com seemed so sketchy.
“Their ads are lame,” I told Ruby (some of whose experiences described here are actually those of other friends of mine; I was keen to protect her anonymity). “You’re young and beautiful, and you live in a city! Why are you wasting your time on strangers?” But Ruby grew up in the Midwest and has never managed to shake the conviction that dogged work is correlated to success in all realms, and she was determined to apply herself strenuously to dating on Match.com.
She tweaked her profile and searched countless men’s profiles, exchanging e-mails and meeting up with the most likely-seeming specimens several days a week. At one point, she even sacrificed a weekend morning to have a coffee date with a PhD student in Continental philosophy who didn’t ask her a single question about herself. As the month progressed, her standards became lower. Men who cropped their profile pics at the eyebrows to exclude their retreating hairlines were no longer off limits. But as Ruby went downmarket and reset her “what I’m looking for” settings accordingly, she succeeded only in increasing the quantity of her miserable dates. At the end of the month, she was forced to admit defeat. But we were left wondering how America’s most popular for-fee dating site, which boasts that it attracts 20,000 new members a day, had failed Ruby.
Curious about the site’s matching algorithm, I set up my own profile. Because I already knew what a woman’s experience on Match was like, I posed as a male version of myself. I bumped my height up two inches and used a rather nebbishy-looking Facebook friend’s profile photo as my own.
The resulting persona, SensitiveDude450, was a five-foot-nine-inch, “athletic and toned” 27-year-old Jew with an annual salary in the mid five figures. He liked yoga and cats. And while some women did look at his profile during the month I spent on the site, no one ever sent him an e-mail or even a “wink.”
To be fair, SensitiveDude450 was not exactly putting himself out there. Proffered “mutual matches,” he declined to e-mail them. But these matches, and the “Daily 5” (selected by the site’s “advanced matchmaking service,” which prompts the user to check out the day’s matches and select “yes,” “no,” or “maybe” for each profile), did contain some clues as to how matching works–and I needed clues, because no one who works at Match would speak to me for this review, on the not-unreasonable grounds that the site’s methodology is proprietary. The short answer to the question “How does matching work?” seems to be: “Not the way it pretends to work.”
SensitiveDude had not expressed any preferences when it came to the height, ethnicity, salary, or body type of his potential dates. “Surprise me!” I thought.
But in the first batch of suggested dates, the matches seemed to have less in common with the Dude than with each other. That set included a short 23-year-old Jewish woman with a cute photo of herself at a fancy restaurant and two others of herself on the arms of guys. She liked the Economist but also Us Weekly. Her favorite things included “brunch.” The site said we had been matched because we were both dog lovers, eldest children, and athletic and toned.
The second match, too, the site said, was deemed compatible on the basis of birth order and pet preferences. But I noted that she was also Jewish, also young (24), and also short, at five foot two. “Seats at a Yank’s [sic] game are always a winner with me,” her “About Me” section declared. This assertion was reinforced by a photo of her in a jersey at a baseball game. Another photo showed her posing in panties and a tank top.
I clicked “maybe” when the site asked me to say whether I was interested in her, and then I clicked “maybe” on a couple of the other short young Jews–not wanting to click “yes,” which would have automatically informed the women of my interest. But one woman who’d been shown my profile in her top 5 did click “yes,” so I checked her out.
She was a 24-year-old lab tech at a fertility clinic, with an incoherent, heavily misspelled profile. She loved malls and hated country music, and her profile photograph was an odd shot of her sucking on a straw. The site, seeming desperate to find something we had in common, pointed out, “Like you, she’s never been married!” I looked at my own profile to remind myself that I was no prize, but then I shut my laptop. I was beginning to understand the basis of the distrust I’d felt when Ruby joined Match. It was gross to know that actual men sat there as I’d just done, flipping through photos of women so desperate for their attention that they posted photos of themselves in bathing suits, twisting around to accentuate their butts while delivering soft-porn smiles.
All this is big business. Online dating, according to Forrester Research, produced $957 million in revenue in 2008–making it the third-largest generator of online paid-content revenue, after music downloading and gaming–and is expected to grow another 10 percent annually through 2013. Even (or especially) in the face of economic contraction, Match.com is thriving.
As a man on Match, I had the sense that what I was doing was a kind of online shopping, which makes sense. The site uses the same type of data-mining technique, called latent semantic indexing (LSI), that search engines like Google use to rank the relevance of Web pages.
The trick behind successfully matching people and products–or people and other people, or people and other people who’ve packaged themselves into something like products by means of “profiles”–is math. “You and I don’t imagine four-dimensional spaces, but mathematics and computers can,” says David Jacobs, a vice president at the blogging-platform company SixApart, who’s worked with similar technology in designing social-media sites. “Each additional attribute considered creates an extra dimension in the ‘space’ with which Match.com is looking for matches. The algorithm creates a virtual graph which approximates hundreds or thousands of axes.”
That’s straightforward. But the other half of the trick is not: it has to do with analyzing the way customers browse rather than the rankings and feedback they deliver. It’s the difference between recommending a match for SensitiveDude450 because we’re “both eldest children” and recommending a match because the site knows that users like SensitiveDude click on the profiles of women who make a bit less money, are shorter, and share the same religion.
“Each of those companies invests heavily in R&D to try and find ‘cheats’ [that they use as] a competitive advantage,” says Jacobs. “They can’t ever share details, because they consider it a secret sauce. Also, my guess is that these cheats are not along single vectors, although ethnicity would probably be straightforward to identify as something that people would claim not to care about when, of course, they did.”
By “cheats,” Jacobs doesn’t mean that Match’s developers have automated their insights about who tends to like whom. More likely, the programmers use an algebraic tool called singular-value decomposition, or SVD, which has many applications in statistics. Match.com’s computers are ignorant of the qualities that humans are thinking of when they use terms like religion or body type. Instead, they recognize patterns: SVD assigns values to the likelihood that two users with various combinations of stated preferences and characteristics will think each other a good match.
After Jacobs had filled me in on LSI, it made sense that the explanations Match gave me (“You share a birth month!”) were simplifications. It generated them after it found a match by observing whose profiles I spent the most time reading and whose profiles other users like me have liked, among any number of other factors.
It’s creepy, the idea that a computer can suss out what it is that SensitiveDude really wants–or at least, what he would be looking for if he existed. The only thing that makes it less creepy is that, at least in Ruby’s case, all that predictive technology turned out–over and over again–to be wrong.
More time spent on the site might have paid dividends for Ruby: the site would have gotten to know her better. Lately, though, she has been searching the offline world for matches. This approach has its upside. For starters, you can wait until after you’ve actually met someone to show him what you look like in a bikini.
Emily Gould is a former editor at Gawker.com.
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today