To date, we’ve collected more than ten million judgments on over one hundred thousand faces. On a lazy Saturday afternoon, we finally dumped the data and played around with it.

Aggregating millions of these snap decisions tells us a lot about our own biases in surprising ways.

For example, you might think that 20-year-olds would be judged as most attractive. However, in this data babies are most attractive, with another peak around 26. After a dip from 40-50, attractiveness starts to increase again.

We have far more data on people between 18-40 on our website, which explains the tighter error bars.

Women are judged as much more trustworthy than men, with the lowest scores for adolescent males. Interestingly, there is a large jump in trustworthiness for both men and women between 20 and 30, and between 50 and 60:

Children and old people are judged as more intelligent, with males in their twenties getting the lowest scores.

As men and women get older they are thought to be more and more conservative. It’s interesting that young women are perceived as more liberal than young men, but the gap disappears after 25.

A few more details on the above.

• FaceStat has more female than male users. Users both upload faces and also judge others. Judgments were collected over the last eight months. All faces have at least 100 judgments each.
• We grouped faces by perceived age, one bin per year, and plotted one point for the average value of the attribute. The y-axis is normalized: centered on the average face rating, with tick marks for +/- 1 standard deviation across faces.
• Error bars are 95% confidence intervals, though omitted on small sized bins where they would be extremely large. In some sense they are too large (too conservative), since each age year is treated separately. The line is a loess fit.

Finally, here’s a scatterplot matrix of the attributes. Every pair of attributes has two graph panels. The bottom-left panels are smoothed scatterplots that show the density of faces in that attribute pair’s space. The top-right corrgram panels show Pearson correlations: blue means the two attributes are positively correlated, and red means negatively correlated. Unlike the above graphs, genders are not separated, and values are not normalized.

There’s a story in each panel. Looking at the attributes in the middle, we see that conservativeness, wealth, intelligence, and trustworthiness all seem to go together. Intoxication has lots of red panels: it’s anticorrelated with all of them. Age would go along being correlated with all these things, except that extreme youth gets high intelligence and trustworthiness marks. Attractiveness is more complex too: it sometimes goes down at the extremes. Perceived political moderates look more attractive compared to liberals and conservatives; similarly, you’re hot if you look moderately smart or rich, but hideously high wealth and intelligence are a little less attractive.

(On the age-attractiveness scatterplot, note the “old beautiful people” effect seems to weaken compared to the gender breakdown graphs earlier in this post. Simpson’s Paradox?)

-Lukas and Brendan

Update 7/11: The graph explorer is now part of the main facestat.com site! Check it out: facestat.com/cloud. And stay on FaceStat to help us collect more data :)

But on Sunday, an insane number of people suddenly decided to flock to our site. Let’s extend the previous chart by 2 days, then a little bit of y-axis auto-scaling says it all:

Turns out the giant spike was due to our being featured via a news article on Yahoo.com’s front page!

Of course, we had to frantically rearchitect the system and scale it under this deluge of traffic. You can read the blow-by-blow account of our crazy few days on Lukas’s blog, here.

The web startup community seems pretty interested in the mad scaling issues, so I’ll respond to some of the comments on Lukas’s blog below:

Yes, we’re pretty much using Rails. We actually use an offshoot called Merb — which is a bit more efficient — on top of Thin. We find that a Rails-like platform is invaluable for rapidly prototyping a new site, especially since we started FaceStat as a pure experiment with no idea whether people would like it or not, and with a very different feature set in mind compared to what it later became. And it’s invaluable that Chris on our team is such a Ruby expert :).

However, the high-level platform really doesn’t matter compared to overall architecture: how we use the database (postgres), how much we cache (memcached/merb-cache), how we distribute load, how we deploy new systems (xen/slicehost), etc. It’s hasn’t been trivial since FaceStat is write-heavy and performs fairly complex statistical calculations, and various issues remain. But we are serving many users at nearly 100x our old load, so something must be going right — at least for now!

-Brendan

p.s. Thank you, Google Analytics, for the above charts. Some day when I grow up, I hope I am wise enough to create an equally brilliant data visualization tool.

Some interesting trends (where the first word often occurs with the others):

• Old: dad, wise, jolly, pedophile, grandpa, perv, professor, writer, sour, experienced, dead, matronly, alcoholic
• Smart: business, librarian, azn, graduate, genius, engineer, intent, bookworm
• Oily: wet, sweaty, shiny
• Drunk: drunkard, tipsy, alcoholic, partyboy, scene, stripper, wasted, ditzy, ew
• Young: tween, underage, uninterested, teen, childlike, jailbait, child, tooyoung, highschooler, babyface, kid, boy, innocent, virgin
• Athletic: runner, driven, jock, sporty
• Serious: angry, grumpy, direct, piercing, alert, dedicated, azn, doctor, suave
• Nerd: goodlooking, virgin, slacker, nerdy, goof, geek
• Gay: cowboy, flamboyant, metrosexual, yuppie, feminine, queer, homosexual, out, pissed, dangerous

Other highly concurrent pairs: army & dedicated, fighter & patriotic, skeezy & hairy, mustache & dad, sunglasses & secretive, cougar & milf, naked & creepy, pimp & playa, plastic & fake, tease & sexy, badass & cool.

For more information on how I counted concurrence, in this diagram I would count the tag “tall” occurring 4 times with the word “lanky”, even though the concurrences are all on one line. I converted all words to lowercase.

I sorted the list of tag pairs by a formula: number of concurrences of the pair / number of total occurrences of the first tag. This sort helped to find the interesting pairs.

-Mike

I was looking at some photos of couples on FaceStat and wanted to run a quick and dirty version of this experiment, and including the other FaceStat variables. I photoshopped out the girl from guy/girl photos and reuploaded. The results followed the same trend as the Aberdeen study (although the Turkers are mixed gender). The strongest, most consistent difference for the photoshopped photo was in relationship status (more single, obviously) and in attractiveness (25% of people found them less attractive). Some examples:

Some details and other results:

The guy’s face was clearly identified in the couple photo. I know this because 100% of Turkers classified the person in the couple photo as male.

The single guys were also consistently identified as more conservative and heavier.

-Mike

If you could ask dozens of random people on the internet questions like “Do I look smart?” or “Does this dress make me look fat?”, why wouldn’t you? Today there are no excuses! With the launch of FaceStat you have the power to decide exactly what profile picture is right for the 10 different social networks you’re a part of. You have the power to make sure the photo you’re about to upload to Match.com makes you look attractive and funny. You have the power to make sure you look trustworthy and never divorced before your girlfriend sends that picture to her parents.

A couple weeks ago we took some pictures and asked Turkers a bunch of questions. We thought this was really cool so we built a tool to let anybody see what people had to say about their picture.

You just upload a picture from your hard drive or Facebook, and we send the faces out to Mechanical Turk to be judged. Currently you can get statistics on one photo per day. Browse the latest faces, or even see who looks the most trustworthy.

Add some faces, and see how you stack up!

-Chris

Intoxication was listed as a checkbox. Averaging the guesses for each photo gives us an intoxication rating from 0 (sober) to 1 (smashed). In the histogram, you’ll notice the familiar power law distribution of drunkenness:

Finally, a cautionary graph. We found a clear correlation between perceived intoxication and perceived intelligence:

Intelligence scale: 0 = “doofus”, 1 = “dull”, 2 = “average”, 3 = “bright”, 4 = “genius”

-Mike

You can see the photos ordered on three axes here:

More details after the jump:

There was mostly consensus from Turkers on guessing ethnicity. Next time we might compare Turker opinion against the profiles’ self-reported ethnicity.

The median error on guessing age was 2 years too young – which makes sense, as the photos were likely taken a year or two ago. Some of the larger negative errors appear to be on photos of 30- and 40-year-olds from high school. Here is a histogram for the error on the guesses:

Unsurprisingly, perceived intelligence correlates with perceived age.

As you can tell we are mostly having fun with this set of questions, but we have some good ideas on how to refine the experiment. If you’re interested in being analyzed in the next round, join the Dolores Labs Facebook group.

Look for some intoxication analysis in the next post.

- Mike