Chris McKinlay - Data Mining for Dates

 

So, it's the summer of 2012, and I'm living in my thesis cubicle. [audience chuckles] I'm in the sixth year of a PhD program in math at UCLA. At that period in my life, I was going through this monastic phase, and I had a little foam pad that I would roll out under my desk to sleep, and I would go to the UCLA gym to take a shower. Every night, I would log on to this supercomputer called Yellowstone, which is at the National Center for Atmospheric Research. At the time, it was in the top 20 powerful supercomputers in the world. And so, I could only use it at night. 

 

And so, every night I would log on and I would run these distributed data mining algorithms that I was working on for my dissertation. It was pretty quiet in the cubicles at night. And so, when my code was compiling or running on the cluster, I'd have a lot of downtime and I'd kill it on Reddit or OkCupid. [audience laughter] And one night, I'm reading OkCupid's blog and I come across this post, this line in a post, and it says, "OkCupid is no more responsible for your match percentages than Microsoft Excel is responsible for your net worth." [audience laughter] I took that as an invitation to program my way into a relationship. [audience laughter] I'd been questioning my lifestyle at the time [audience laughter] and wondering if maybe it was time to try and find someone and start a relationship. So, I decided that I was going to reverse engineer OkCupid's match algorithm. [audience laughter] 

 

In case you don't know what OkCupid is, it's the largest free dating website in the world. It's got about 14 million active monthly users. And between any two of those users, it generates a number between 0 and 100 called a match percentage. And everything in the site, all the activity, all of the visibility is based on these match percentages. Your match percentage with any other person is determined on the basis of your answers to these multiple-choice questions called match questions. There's about a half million of them. They're all user submitted, cover all walks of life, dating, drugs, sex, religion, lifestyle. 

 

And so, I'm looking at these questions and I ask myself a question. "Well, okay, what happens when everyone goes online and answers these? Do their answers uniformly percolate throughout the space the way that they might if we all flipped coins to determine our answers, or do they tend to clump up somehow around commonly held belief systems or something like that? And if so, how much?” So, the more I thought about that, the more compelling the idea seemed. I started using all of my supercomputing time to download and analyze OkCupid match data. [audience laughter] 

 

I discovered that when people in Los Angeles go online and answer match questions, they don't do it uniformly at random. Instead, they cluster rather tightly into seven different groups. And so, I started checking out these different groups and I found one that looked datable. [audience laughter] So, I wrote some combinatorial optimization software to figure out which of these half million questions could I answer truthfully, but still maintain the highest possible match percentage with every single person in the cluster. [audience laughter] 

 

So, I wrote that and I ran that code on this supercomputer for a while, [audience laughter] and it spit back the right questions to use. I entered those into my profile, and that's when things blew up. [audience laughter] All of a sudden, I became the top match for over 30,000 women in Los Angeles. [audience laughter] [audience cheers and applause] 

 

So, that meant that anytime any one of these women went on OkCupid, I'd be there at the top. [audience laughter] Any search they ran, anytime they looked at their what's new feed, there I was. [audience laughter] But I'm still this ratty grad student with a bad profile, and I'm getting hundreds of profile views a day. I got excited and decided to write a whole bunch more software to optimize [audience laughter] down to the very last word and last pixel, like the ideal profile for myself to convert these page views into messages.

 

So, after doing that, I started getting about 88 unsolicited messages per week, which, to put in perspective for a straight male and OkCupid, the median number of unsolicited messages per week is zero. [audience laughter] So, that's how for three months during the summer of 2012, I became the most popular of the seven million male profiles on OkCupid. When it was happening, I was like, "Oh my God, this is amazing. This is going to be like Amazon Prime for my dating life." [audience laughter] I remember being worried like, "Is OkCupid going to notice? [audience laughter] Are they going to shut me down? How long do I have?" [audience laughter] 

 

So, I decided I needed an equally hardcore second phase to this project, and so I promised myself I would go on an average of one date per day until they either shut it down or I met someone worth shutting it down for. I still remember the first date that I went on after deciding to do that. So, I unfold myself from my cubicle, and I go to the gym, and shower. I had to leave early, because I knew I was going to have to jump start my car. [audience laughter] So, I get a jump and I drive to West Hollywood and I meet this entertainment lawyer. She's driving a Mercedes convertible, [audience laughter] and she wants to go to this expensive restaurant. I'm like, "All right, this is not even worth doing. Is it dead? This is a total waste of time and money." But whatever, I'll go. [audience laughter]

 

And over dinner, I'm thinking, why is she even doing this? She looks at me quite seriously and asks me, "You know, I've been on OkCupid since 2006, and I've never seen a 100% match before. Do you think we're soulmates?" [audience laughter] I got really freaked out [audience laughter] and I thought, wow, I'm playing with fire. I have to be very careful. [audience laughter] And so, that whole month, I just kept going on these dates. Dinners, concerts, hikes, one funeral. [audience laughter] I was trying to have these romantic, storybook first dates, and it was just killing me. 

 

I put over a thousand miles on my car, I was burning through time and money, and it was clearly not scalable the way I wanted it to be. [audience laughter] And so, I asked myself a second question like, "Well, all right, what is the smallest socially acceptable interaction I can have around about three minutes of FaceTime?” Because that's all I really need to know if I want to continue, and I haven't been able to write software to really tell me that information. So, I decided on a small cup of decaf coffee, preferably drunk, standing up. [audience laughter] I found these four coffee shops in Westwood, close to UCLA. They were all kitty corner to one another. And so, I would just do one coffee date each hour on the hour [audience laughter] and work my way around. 

 

At that point, I'd started to notice some interesting similarities between these dates. These were people who'd answered several hundred of these multiple-choice questions with a high degree of statistical similarity. But the common frame of the coffee date really brought it out. So, similar conversations, similar affectations. At one point, eight out of nine people in a row ordered the same coffee drink. [audience chuckles] And so, I think partly as a result of the practice and also my, I think, natural inclination to try and optimize things, I got very good at these dates. [audience laughter] 

 

And in the course of six weeks or so of doing that, I took the conversion rate from first date to email asking for second date from 50% to 60 to 70 to 82%. [audience chuckles] But the problem was I wasn't going on many second dates, because at any given time I had six or seven first dates already lined up. [audience laughter] And so, I was saying no to people that just a few months ago I would have been completely ecstatic to go on dates with. I was completely desensitized. And yet, these women were people who matched very highly with me, and so they were bringing their A game to these dates, they had expectations. I was, I think, disappointing those expectations, and it became very hard for me to do that. I felt guilty. These didn't feel romantic in any way. 

 

So, in order to avoid writing these hard emails, I began to intentionally tank dates. So, at first, I would just tell them the truth, "Hey, I've got three other dates lined up today." [audience laughter] And that didn't work. [audience laughter] That cut the second date email percentage by about 12%, but not more than that. So, I started to go to more and more extreme measures. So, one thing I would do is in the middle of the date, if I decided, "Okay, no good," I would go to the bathroom and I would put on eyeliner, [audience laughter] and then I would come back, and she'd be there and she'd say, "Did you just put on eyeliner?" [audience laughter] And I would completely deny it. [audience laughter] Completely, awkwardly deny putting on eyeliner. 

 

That was much more effective. [audience laughter] But clearly, something had gotten twisted. [audience laughter] I began to question my motives. "Are you really trying to find someone, or are you just trying to optimize some game, or are you more into the sociological side, or is it the performance art side?" [audience chuckles] I thought about gathering more data, refactoring the code. I thought very seriously about shutting it down. But I didn't, because at that point, I was going viral, I guess, on OkCupid. I was getting placed in front of enough eyeballs that I was getting messages from Argentina, from Australia, from all over LA, from people who didn't belong to this sort of cluster that I'd optimized for. And so, every once in a while, on these dates, there'd just be a total surprise. I loved that. 

 

So, I remember on date number 88, I met this woman named Christine. She was like, "Hey, you know, your profile says 'replies very selectively.' So, I decided to amp up my profile for you, and put these pictures on, and tweak the language a little bit." I could tell that she was really different, and I just melted and said, "Oh my God. Well, I have a question for you. I basically did the same thing. Actually, I did much worse. I used some topic mining algorithms to plot all these things, and then I reverse engineered the match algorithm, and I've been going on an average of 1.1 dates per day for the last 80 days, and I'm getting really tired and I'm confused, but you seem really cool. Is that like twisted?" [audience laughter] 

 

She took a beat, and she said, "No. You know, that's actually what it's like being a woman dating online. [audience chuckles] You have all these people writing you, you don't really know. If it's about you, it's confusing." We had a really good 20-minute conversation about the politics of gender and data. It was a really good date. It felt like a good date. Anyway, that was a couple years ago. Since graduated and moved out of my cubicle and moved on with my life, but I occasionally tell this story. When I tell it, people often express this like buyer's remorse about their relationships, "Oh my God, I met my significant other in a bar or through friends. Do you think I stopped dating too soon?" [audience laughter] They want to know, "Well, did it work out for you? I mean, did you meet someone or did it all fizzle [chuckles] despite all that work you put into it?" Fortunately for me, it did work. Christine and I are engaged. [audience applause] 

 

But it's not because I hacked OkCupid. [audience chuckles] I would say we're engaged because we work pretty hard at our relationship. And relationships, at least good ones, aren't the kinds of things that you can get from Amazon Prime. [audience chuckles] But I usually also add when I respond to these people that I'm fairly certain I conducted a thorough search. [audience laughter]