The simplest thing would be to just give people that list of tips from the beginning of the article. That would help them find the best photo a lot easier than this. There's also a lot of studies from ok cupid on profile picture optimization.
Second, how many people get enough matches to get any statistically useful information? The people that need this the most are the least likely to benefit.
Third, epsilon greedy isn't the best algorithm to use. Thompson sampling is theoretically optimal, and pretty simple. Here is a cool simulation of it in your browser. I find it fun to play with (sadly the graphs no longer render properly on this archived version. I have no idea why. But it used to display a beta distribution for each bandit, and you could see it shrink over time.): http://web.archive.org/web/20160310000407/https://e76d6ebf22...
Lastly it would be interesting to train a neural network to predict how good a profile picture is. This has sort of been done by Karpathy here: http://karpathy.github.io/2015/10/25/selfie/ Tindr has an enormous amount of data that could be used.
I believe they also differentiate between immediate right-swipe, and profile view before right-swipe. Sometimes my top picture changes and I haven't had any matches in a while.
Or they're just fucking with me to keep my attention.
I don't just want any right swipe. I want a right swipe from the type of person I want to meet. The image tinder claimed was best is IMO not representative of me. I have it in my set of images to show I "clean up well" but if someone chooses me for that picture I feel like they'll be disappointed and most likely not my type therefore I don't have it as my first picture.
Then you shouldn't use Tinder. Okcupid was trying to do smart matching between people but apparently everyone is lazy nowadays and prefers swiping on photos rather than reading.
Welcome to 2017
Tinder calculates a rating for each user (ELO) using the data on right swipes (and other things). Then it only shows your profile to users with a similar rating. (Many users have documented this feauture)
Tinder wants the most influential users to recommend the app to their friends (on social media and in real life). Influential users are typically more desirable and have a higher rating.
So you should take all the right swipes you can get.
"Smart Photos" are fascinating. I travel frequently, and depending on the country that I'm in, within 24 hours, a new smart photo will filter to the top.
I'd love to see research into this - revealing which cultures / countries / societies place importance on which elements.
For example:
- France, Italy: Suit and tie selfie springs to the top, constantly. Means that french / Italian women are more looks focused?
- Australia, Spain: Outdoor / adventure photo, straight to the top. Probably no surprise, Australians tend to be more outdoorsy.
- Asia: Presentable smart casual goes to the top. Looking more for 'cute' or 'presentable' ?
It'd be very interesting to apply the same profile to all countries to see what floated to the top.
I love Machine Learning. I think it is one of the biggest breakthroughs of this decade.
The Big Data + Computing power recipe is something that I thought will bring sweeping changes to society. But with Amazon's Echo look, and now this, I don't know if I'm ok with where this is going.
I thought it'll help us in fields like architecture, physics modelling, medicine, but they all seem to be on paper. OTOH, uses like this seem to be the ones that actually come to market and make the best use of the Machine Learning talent.
Maybe I'm being cynical, but please someone prove me wrong. Am I missing Machine Learning being used in new applications in the market, or is the situation as I see it is?
I'm seeing a more general trend which I find equally worrying:
Software is eating the world, and the best tech is currently produced by the likes of Facebook, Google or Amazon.
I'm fine with this, but I'm seeing governments and other "serious" businesses like health, physics... seriously lagging behind.
I don't know what to do with this assumption, though...
Google will run the world? I'm saying it's too good to be true.
First of all there is no machine learning used in the article.
Second, machine learning IS helping in places like physics and medicine. There's currently a one million dollar prize by google for an algorithm that can detect cancer (actually it just ended.) And there have been a number of similar contests in the past. Just look here for all sorts of places machine learning has been applied: https://www.kaggle.com/competitions?sortBy=deadline&group=al...
People have been applying statistics and AI to medicine since the 80's. The biggest problem is medicine and similar fields, is it's terribly conservative. It's very difficult to get new technology to market. Whereas tindr doesn't have any regulations or bureaucracy to deal with. They can just push new technology to millions of users overnight.
I used the term machine learning in my first sentence, but I was actually referring to data + computing power as I did in my second paragraph. Even if they just use OpenCV, that still falls into the realm of using computers to solve problems that are only recently being solved due to the computing power being recently available.
>The biggest problem is medicine and similar fields, is it's terribly conservative. It's very difficult to get new technology to market. Whereas tindr doesn't have any regulations or bureaucracy to deal with. They can just push new technology to millions of users overnight.
It's things like this that bring machine learning to the masses, which also typically lowers costs and makes it more available for niche uses.
It's a lot easier and more forgiving to use machine learning to pick what picture or outfit you look best in than it is to use it in medicine or architecture. As the technology matures, you will see it move in to those sectors very rapidly.
It has to pay the bills. There is money to be made from the Echo Look so it gets developed first. For things that will help society but aren't profitable we will have to look to other funding sources.
Even though this is better than the lack of it, I find it weird that they're not using Contextual MAB and instead just went with the MAB. You've got the preference of your users anyhow, why not show the picture first that will perform best according to that specific user?
I'm guessing the reason was that this was born out of a hackathon. E-greedy is likely the easiest algorithm to implement for a MAB problem like this. Given that they mention there's an API to specify sort order (assuming it was on a per-view basis), the specific algorithm could be swapped out for something smarter.
I think a contextual bandit algorithm would be interesting in this case, although I wonder if there would be enough data behind photo swipes to back each context meaningfully.
An odd caveat of Smart Photos on Tinder is that Tinder doesn't tell app users that their public photo may not be the same as the one that they themselves have assigned. It might be a smart business idea for Tinder to charge for profile-pic analytics.
Huh. I noticed in the profile the pic order changed, which made me think it was a bug until I remembered Tinder is more data-savvy, but I never received a notification.
Sounds like the algorithm doesn't care whether or not you liked somebody. So even for a very indifferent person liking, say 1 in 2 people you're still likely getting a nonoptimal result. And AFAIK the whole point of this app is mutual likes?
I think the algorithm solves the problem of which photo, out of a set of photos, is the best one to show. So nothing has changed regarding the mutuality aspect, only which photo of the user is their best photo based on data collected from past swipes. In other words, they're continuously rotating through a user's photos, trying to find the one that gets the most positive feedback and showing that one more often than the user's other photos.
I think the algorithm would show which photo the "most" users would like is. What if you're someone who loves cats like no other and puts a cat on your shoulder (or wears a cat t-shirt or something). You're probably going to have a lot less "likes" than a "normal" picture of you, but the odds of you getting along might be much lower (I'm sure the average individual likes cat people less than the person who picked a person with cats in their picture).
It seems the shallowness of starting on the footing of what someone looks like alone kind of renders this point moot. If you're trying to identify compatibility based on who a person really is, their profile photo may be the least ideal place to start. If anything, I think this at least gets the user a foot in the door more often.
Interesting post - I'm fascinated by this space myself. I launched a site (http://judg.me) a few years back with a slightly different take on how users perceive your profile photos.
I understand this article is a little dated, but for a follow-up blog post, I'm interested in hearing more about Tinder's internal hackathons. What motivated hosting an internal hackathon? What format was chosen, i.e. duration, criteria for projects, criteria for participants, etc.? How often are they held? Do employees retain any rights for the projects they work on? How is participation encouraged?
I don't know about Tinder, but the (telecom) software development company I work for holds internal hackathons twice a year (one in May, one in November). We've been doing this for about 6 years, with over 150 attendees each time from offices around the UK. We normally have teams of 3-5, with a focus on innovation (e.g. bots, IoT, AI). Each runs for 24 hours from 17:00 on Thursday through to 17:00 on Friday when pitching/demos start (2 minutes each) followed by prizes and food/drinks/party. Any work produced belongs to the company (bear in mind this is mostly during work time) although we look at most innovative ideas for patent angles, and we get bonuses for patents that get filed as a result. Prizes are relatively token, but prestigious. Participation hasn't actively had to be encouraged - most engineers who are available attend. The main motivation for the hackathons are innovation and morale (everyone seems to enjoy them), but other benefits are education (learning new skills, languages, APIs, etc.), working with new people/teams and personal development (e.g. it's not unusual for new employees to take on technical lead roles that they wouldn't have normally).
We also hold "vacathons", which are hackathons just for the interns in the summer. There are about 50 attendees, and because they haven't always done too much software development beforehand, these are 3 days long and people are less likely to work through the night (although we lay on food in the evening, and people often stay until 10ish). The theme for these is more general - anything loosely to do with communication is accepted. As with the full hackathons, teams form and generate ideas themselves, but we also assign a mentor (a full-time employee) to each team to help them with any technical problems as well as planning/working as a team.
The problem is that swipers are not uniform. You might strongly select for cat people with a photo of you with a cat, but negatively with hotties. So unless you can boost your ELO into the next strata with the aid of cat people, beware. And even then the stats of each stratum may differ, so you could oscillate up and down.
This reads like it was written by a person on the hackathon team, who truly believed in his project. I sincerely doubt those two things were all the perks provided.
To me, he's just trying to paint a picture filled with drive. One where he planned on working overnight on something in which he believed. No need to insinuate tinder doesn't care about its engineers or draw cynical conclusions.
No need to defend a multi billion corporation getting its engineers to do unpaid work either (unpaid overtime I assume), but hey there's a sofa to sleep on - fk that.
Isn't this false advertising? If someone doesn't have the wherewithal to know how to choose a picture that best represents themselves, then I don't want to talk to the actual person.
Stop trying to optimize everything for my engagement and fix other people's stupidity.
The next thing you'll be telling me is you're automatically removing blemishes and making people look more attractive.
Tinder is a glorified personal ad. Stop thinking you're some kind of advanced tech shop worthy of such algorithms with fancy-sounding curves.
Second, how many people get enough matches to get any statistically useful information? The people that need this the most are the least likely to benefit.
Third, epsilon greedy isn't the best algorithm to use. Thompson sampling is theoretically optimal, and pretty simple. Here is a cool simulation of it in your browser. I find it fun to play with (sadly the graphs no longer render properly on this archived version. I have no idea why. But it used to display a beta distribution for each bandit, and you could see it shrink over time.): http://web.archive.org/web/20160310000407/https://e76d6ebf22...
Lastly it would be interesting to train a neural network to predict how good a profile picture is. This has sort of been done by Karpathy here: http://karpathy.github.io/2015/10/25/selfie/ Tindr has an enormous amount of data that could be used.