How Tinder creates better matches using AWS image recognition technology
- 08 December, 2018 09:29
Popular dating app Tinder is using image recognition technology from Amazon Web Services (AWS) to power its matching algorithm for premium users.
Speaking during AWS re:Invent in Las Vegas last week, Tom Jacques, VP of engineering at Tinder explained how it is using the deep learning-powered AWS Rekognition service to identify user's key traits by mining the 10 billion photos they upload daily.
"The challenges we face are in understanding who members want to see, who they match with, who will chat, what content can we show you and how do we best present it to you," Jacques outlined.
Tinder ingests 40TBs of data a day into its analytics and ML systems to power matches, which are underpinned by AWS cloud services.
Jacques says that Tinder knows from its data that the primary driver for who you match is photos. "We see it in the data: the more pictures you have, the higher likelihood of success to match."
When a user joins Tinder they typically post a set of pictures of themselves and a short written bio, however Jacques says an increasing number of users are foregoing the bio altogether, meaning Tinder needed to find a way to mine those images for data that could power its recommendations.
Rekognition allows Tinder to automatically tag these billions of photos with personality markers, like a person with a guitar as a musician or 'creative', or someone in climbing gear as 'adventurous' or 'outdoorsy'.
Tinder uses these tags to enrich their user profiles, alongside structured data such as education and job information, and unstructured raw text data.
Then, under the covers, Tinder "extracts all of this information and feed it into our features store, which is a unified service that allows us to manage online, streaming and batch processing. We take this information and feed into our tagging system to work out what we highlight for each profile."
In short, Rekognition provides Tinder with a way to "access what is inside these photos in a scalable way, that's accurate and meets our privacy and security needs," Jacques said.
"It provides not only cloud scalability that can handle the billions of images we have but also powerful features that our experts and data scientists can leverage to create sophisticated models to help solve Tinder's complex problems at scale," he added. "Privacy is also important to us and Rekognition gives us separate APIs to provide control and allow us to access only the features we want. By building on top of Rekognition we are able to more than double the tag coverage."
Premium users of Tinder also get access to a Top Picks feature. Launched in September, this provides Gold users - the most expensive bracket at around £12 a month - with a curated feed of "high quality potential matches".
All Tinder users receive one free Top Pick a day, but Gold subscribers can tap a diamond icon at any time for a set of Top Picks, which is refreshed daily.
"When it comes to serving this when a member wants their Top Picks we query our recommendation cluster, the same underlying technology that powers our core recognitions, but looking at the outcomes users are trying to achieve and to provide really personalised, high quality matches," Jacques explained.
"Top picks has shown a great increase in engagement compared to our core recommendations, and beyond that, when we see these tags on profiles we see a further 20 percent lift." Jacques said.
Looking forward, Jacques says he is "really excited to take advantage of some of the recent features that have come out [from AWS], to enhance the model accuracy, added hierarchical data to better categorise and cluster content, and bounding boxes to not only understand what objects are in photos but where they are and how they are being interacted with. We can use this to get really deep into what is going on in our members lives and provide better services to them."
Rekognition is available off the shelf and is charged at US$1 for the first 1 million images processed per month, $0.80 for the next 9 million, $0.60 for the next 90 million and $0.40 for over 100 million.