0:00:17.903,0:00:19.617
Hello, my name is Christian Rudder,

0:00:19.641,0:00:21.850
and I was one of the founders of OkCupid.

0:00:21.874,0:00:24.792
It's now one of the biggest[br]dating sites in the United States.

0:00:24.816,0:00:27.207
Like most everyone at the site,[br]I was a math major,

0:00:27.231,0:00:30.671
As you may expect, we're known[br]for the analytic approach we take to love.

0:00:30.695,0:00:32.333
We call it our matching algorithm.

0:00:32.357,0:00:34.945
Basically, OkCupid's matching[br]algorithm helps us decide

0:00:34.969,0:00:36.845
whether two people should go on a date.

0:00:36.869,0:00:38.741
We built our entire business around it.

0:00:38.765,0:00:40.725
Now, algorithm is a fancy word,

0:00:40.749,0:00:43.234
and people like to drop it[br]like it's this big thing.

0:00:43.258,0:00:45.546
But really, an algorithm[br]is just a systematic,

0:00:45.570,0:00:47.793
step-by-step way to solve a problem.

0:00:47.817,0:00:49.994
It doesn't have to be fancy at all.

0:00:50.018,0:00:51.169
Here in this lesson,

0:00:51.193,0:00:54.201
I'm going to explain how we arrived[br]at our particular algorithm,

0:00:54.225,0:00:55.636
so you can see how it's done.

0:00:55.660,0:00:57.594
Now, why are algorithms even important?

0:00:57.618,0:00:59.198
Why does this lesson even exist?

0:00:59.222,0:01:02.642
Well, notice one very significant[br]phrase I used above:

0:01:02.666,0:01:05.005
they are a step-by-step[br]way to solve a problem,

0:01:05.029,0:01:08.447
and as you probably know, computers[br]excel at step-by-step processes.

0:01:08.471,0:01:10.060
A computer without an algorithm

0:01:10.084,0:01:12.808
is basically an expensive paperweight.

0:01:12.832,0:01:15.821
And since computers are such[br]a pervasive part of everyday life,

0:01:15.845,0:01:17.392
algorithms are everywhere.

0:01:18.590,0:01:21.787
The math behind OkCupid's matching[br]algorithm is surprisingly simple.

0:01:21.811,0:01:25.813
It's just some addition, multiplication,[br]a little bit of square roots.

0:01:25.837,0:01:27.527
The tricky part in designing it

0:01:27.551,0:01:30.116
was figuring out how to take[br]something mysterious,

0:01:30.140,0:01:31.290
human attraction,

0:01:31.314,0:01:34.098
and break it into components[br]that a computer can work with.

0:01:34.122,0:01:36.675
The first thing we needed[br]to match people up was data,

0:01:36.699,0:01:38.691
something for the algorithm to work with.

0:01:38.715,0:01:41.873
The best way to get data quickly[br]from people is to just ask for it.

0:01:41.897,0:01:44.624
So we decided that OkCupid[br]should ask users questions,

0:01:44.648,0:01:47.005
stuff like, "Do you want[br]to have kids one day?"

0:01:47.029,0:01:48.787
"How often do you brush your teeth?"

0:01:48.811,0:01:50.203
"Do you like scary movies?"

0:01:50.675,0:01:52.752
And big stuff like,[br]"Do you believe in God?"

0:01:53.843,0:01:56.907
Now, a lot of the questions[br]are good for matching like with like,

0:01:56.931,0:01:59.087
that is, when both people[br]answer the same way.

0:01:59.111,0:02:01.659
For example, two people[br]who are both into scary movies

0:02:01.683,0:02:05.004
are probably a better match[br]than one person who is and one who isn't.

0:02:05.028,0:02:06.521
But what about a question like,

0:02:06.545,0:02:08.607
"Do you like to be[br]the center of attention?"

0:02:08.631,0:02:11.259
If both people in a relationship[br]are saying yes to this,

0:02:11.283,0:02:13.376
they're going to have massive problems.

0:02:13.400,0:02:14.645
We realized this early on,

0:02:14.669,0:02:17.938
and so we decided we needed[br]a bit more data from each question.

0:02:17.962,0:02:20.725
We had to ask people to specify[br]not only their own answer,

0:02:20.749,0:02:23.014
but the answer they wanted[br]from someone else.

0:02:23.038,0:02:24.539
That worked really well.

0:02:24.563,0:02:26.167
But we needed one more dimension.

0:02:26.191,0:02:28.834
Some questions tell you more[br]about a person than others.

0:02:28.858,0:02:32.253
For example, a question[br]about politics, something like,

0:02:32.277,0:02:34.565
"Which is worse:[br]book burning or flag burning?"

0:02:34.589,0:02:37.399
might reveal more about someone[br]than their taste in movies.

0:02:37.423,0:02:40.042
And it doesn't make sense[br]to weigh all things equally,

0:02:40.066,0:02:41.662
so we added one final data point.

0:02:41.686,0:02:43.710
For everything that OkCupid asks you,

0:02:43.734,0:02:46.563
you have a chance to tell us[br]the role it plays in your life.

0:02:46.587,0:02:48.906
And this ranges[br]from irrelevant to mandatory.

0:02:49.446,0:02:52.668
So now, for every question,[br]we have three things for our algorithm:

0:02:52.692,0:02:54.044
first, your answer;

0:02:54.617,0:02:58.757
second, how you want someone else --[br]your potential match -- to answer;

0:02:58.781,0:03:01.569
and third, how important[br]the question is to you at all.

0:03:02.710,0:03:03.962
With all this information,

0:03:03.986,0:03:07.104
OkCupid can figure out[br]how well two people will get along.

0:03:07.128,0:03:10.134
The algorithm crunches the numbers[br]and gives us a result.

0:03:10.158,0:03:11.310
As a practical example,

0:03:11.334,0:03:13.859
let's look at how we'd match you[br]with another person.

0:03:13.883,0:03:15.072
Let's call him "B."

0:03:16.023,0:03:19.505
Your match percentage with B is based[br]on questions you've both answered.

0:03:19.529,0:03:21.954
Let's call that set[br]of common questions "s."

0:03:22.559,0:03:24.908
As a very simple example,[br]we use a small set "s"

0:03:24.932,0:03:26.573
with just two questions in common,

0:03:26.597,0:03:28.425
and compute a match from that.

0:03:28.449,0:03:30.120
Here are our two example questions.

0:03:30.144,0:03:32.525
The first one, let's say, is,[br]"How messy are you?"

0:03:32.549,0:03:34.645
And the answer possibilities are:

0:03:34.669,0:03:38.030
very messy, average and very organized.

0:03:38.054,0:03:40.114
And let's say you answered[br]"very organized,"

0:03:40.138,0:03:42.898
and you'd like someone else[br]to answer "very organized,"

0:03:42.922,0:03:45.178
and the question is very important to you.

0:03:45.202,0:03:46.694
Basically, you're a neat freak.

0:03:46.718,0:03:49.586
You're neat, you want someone else[br]to be neat, and that's it.

0:03:49.610,0:03:51.625
And let's say B is a little bit different.

0:03:51.649,0:03:53.688
He answered "very organized" for himself,

0:03:53.712,0:03:56.719
but "average" is OK with him[br]as an answer from someone else,

0:03:56.743,0:03:59.145
and the question is only[br]a little important to him.

0:03:59.169,0:04:02.062
Let's look at the second question,[br]from our previous example:

0:04:02.086,0:04:04.142
"Do you like to be[br]the center of attention?"

0:04:04.166,0:04:05.680
The answers are "yes" and "no."

0:04:05.704,0:04:08.699
You've answered "no," you want[br]someone else to answer "no,"

0:04:08.723,0:04:11.114
and the question is only[br]a little important to you.

0:04:11.138,0:04:12.759
Now B, he's answered "yes."

0:04:12.783,0:04:14.559
He wants someone else to answer "no,"

0:04:14.583,0:04:16.857
because he wants the spotlight on him,

0:04:16.881,0:04:19.311
and the question is somewhat[br]important to him.

0:04:19.335,0:04:21.334
So, let's try to compute all of this.

0:04:21.972,0:04:24.475
Our first step is, since we use[br]computers to do this,

0:04:24.499,0:04:26.366
we need to assign numerical values

0:04:26.390,0:04:29.017
to ideas like "somewhat[br]important" and "very important,"

0:04:29.041,0:04:31.252
because computers need[br]everything in numbers.

0:04:31.276,0:04:33.679
We at OkCupid decided[br]on the following scale:

0:04:33.703,0:04:35.649
"Irrelevant" is worth 0.

0:04:36.173,0:04:38.062
"A little important" is worth 1.

0:04:38.538,0:04:40.347
"Somewhat important" is worth 10.

0:04:40.831,0:04:42.585
"Very important" is 50.

0:04:42.609,0:04:46.221
And "absolutely mandatory" is 250.

0:04:46.245,0:04:48.876
Next, the algorithm makes[br]two simple calculations.

0:04:48.900,0:04:52.146
The first is: How much did[br]B's answers satisfy you?

0:04:52.170,0:04:55.963
That is, how many possible points[br]did B score on your scale?

0:04:55.987,0:04:59.199
Well, you indicated that B's answer[br]to the first question,

0:04:59.223,0:05:00.389
about messiness,

0:05:00.413,0:05:01.763
was very important to you.

0:05:01.787,0:05:04.017
It's worth 50 points and B got that right.

0:05:04.375,0:05:06.112
The second question is worth only 1,

0:05:06.136,0:05:08.414
because you said[br]it was only a little important.

0:05:08.438,0:05:09.635
B got that wrong,

0:05:09.659,0:05:12.441
so B's answers were 50[br]out of 51 possible points.

0:05:12.465,0:05:15.073
That's 98% satisfactory. Pretty good.

0:05:15.097,0:05:19.046
The second question the algorithm[br]looks at is: How much did you satisfy B?

0:05:19.070,0:05:22.329
Well, B placed 1 point on your answer[br]to the messiness question

0:05:22.353,0:05:24.306
and 10 on your answer to the second.

0:05:24.745,0:05:28.132
Of those 11, that's 1 plus 10,[br]you earned 10 --

0:05:28.156,0:05:30.751
you guys satisfied each other[br]on the second question.

0:05:30.775,0:05:35.017
So your answers were 10 out of 11[br]equals 91 percent satisfactory to B.

0:05:35.041,0:05:36.192
That's not bad.

0:05:36.216,0:05:38.723
The final step is to take[br]these two match percentages

0:05:38.747,0:05:40.613
and get one number for the both of you.

0:05:40.637,0:05:43.248
To do this, the algorithm[br]multiplies your scores,

0:05:43.272,0:05:44.937
then takes the nth root,

0:05:44.961,0:05:47.144
where "n" is the number of questions.

0:05:47.168,0:05:49.998
Because s, which is the number[br]of questions in this sample,

0:05:50.022,0:05:51.863
is only 2,

0:05:51.887,0:05:55.552
we have: match percentage[br]equals the square root

0:05:55.576,0:05:58.472
of 98 percent times 91 percent.

0:05:58.496,0:06:00.280
That equals 94 percent.

0:06:00.304,0:06:03.508
That 94 percent is your match[br]percentage with B.

0:06:03.532,0:06:06.775
It's a mathematical expression[br]of how happy you'd be with each other,

0:06:06.799,0:06:07.982
based on what we know.

0:06:08.006,0:06:09.792
Now, why does the algorithm multiply,

0:06:09.816,0:06:12.585
as opposed to, say, average[br]the two match scores together,

0:06:12.609,0:06:14.279
and do the square-root business?

0:06:14.303,0:06:16.832
In general, this formula[br]is called the geometric mean.

0:06:16.856,0:06:19.483
It's a great way to combine[br]values that have wide ranges

0:06:19.507,0:06:21.422
and represent very different properties.

0:06:21.446,0:06:23.859
In other words, it's perfect[br]for romantic matching.

0:06:23.883,0:06:27.130
You've got wide ranges and you've got[br]tons of different data points,

0:06:27.154,0:06:30.592
like I said, about movies, politics,[br]religion -- everything.

0:06:30.616,0:06:32.454
Intuitively, too, this makes sense.

0:06:32.478,0:06:35.253
Two people satisfying[br]each other 50 percent

0:06:35.277,0:06:39.229
should be a better match[br]than two others who satisfy 0 and 100,

0:06:39.253,0:06:41.067
because affection needs to be mutual.

0:06:41.091,0:06:43.582
After adding a little correction[br]for margin of error,

0:06:43.606,0:06:46.177
in the case where we have[br]a small number of questions,

0:06:46.201,0:06:47.518
like we do in this example,

0:06:47.542,0:06:48.714
we're good to go.

0:06:48.738,0:06:50.650
Any time OkCupid matches two people,

0:06:50.674,0:06:52.706
it goes through the steps[br]we just outlined.

0:06:52.730,0:06:54.999
First it collects data about your answers,

0:06:55.023,0:06:58.008
then it compares your choices[br]and preferences to other people's

0:06:58.032,0:06:59.999
in simple, mathematical ways.

0:07:00.023,0:07:02.946
This, the ability to take[br]real-world phenomena

0:07:02.970,0:07:05.385
and make them something[br]a microchip can understand,

0:07:05.409,0:07:08.686
is, I think, the most important skill[br]anyone can have these days.

0:07:08.710,0:07:11.133
Like you use sentences[br]to tell a story to a person,

0:07:11.157,0:07:13.641
you use algorithms[br]to tell a story to a computer.

0:07:14.349,0:07:17.382
If you learn the language,[br]you can go out and tell your stories.

0:07:17.406,0:07:19.159
I hope this will help you do that.