Pinterest built machine learning 'Pinnability' to surface more relevant content

Get ready for more content in line with your interest (or at least pins) if a new homegrown effort from Pinterest works out.


With an estimated valuation of $11 billion and counting, it's high time for Pinterest to prove its worth.

One seemingly easy (albeit always easier said than done) way would be to ensure surfacing the right content to the right users based on past habits, which in turn should fuel the digital scrapbooking service's e-commerce efforts.

Such is the problem to solve for other top social networking platforms like Facebook and Twitter, but even so for digital marketplaces like Etsy and Amazon.

In its true do-it-yourself spirit, Pinterest went with a homegrown effort on this one, unveiling "Pinnability" on Friday.

With already more than 30 billion Pins flooding user boards and feeds, Pinterest is aiming to make more sense out of all that through machine learning.

Pinnability is the collective code name for the advanced machine learning models in development to improve the home feed.

According to Pinterest, the idea is the more people Pin, the better Pinterest can be for each user, in both surfacing relevant content and proving to be a more fruitful experience overall.


Yunsong Guo, a software engineer at Pinterest, explained further in a blog post on Friday how Pinnability's foundation and initial training data starts with a user (or "Pinner's") actions with home feed Pins over time.

"Our unique data set contains abundant human-curated content, so that Pin, board and user dynamics provide informative features for accurate Pinnability prediction," Guo wrote.

Still in its infancy, Pinnability has already produced "significant boosts in Pinner engagement," Guo touted, including a 20 percent uptick in repinning from the home feed. Thus, Pinnability's machine learning models will be applied to other projects on the platform, he added.

Deep learning and machine intelligence have proven to be hot topics this week.

Nvidia CEO and co-founder Jen-Hsun Huang unveiled several new technologies in line with its deep learning strategy amid the GPU Technology Conference in Silicon Valley on Tuesday.

Huang put the spotlight on the Pascal GPU series, which promise to speed up deep learning applications tenfold compared to Nvidia's current-generation Maxwell processors. Pascal's objective is "mixed precision" computing, said to offer greater accuracy in getting computers to teach themselves.

But the Titan X GPU series, touted for cutting-edge mobile gaming and deep learning alike, might reach a broader audience -- and by extension -- spread more awareness about deep learning in general.

To spur development and deployment, Nvidia also developed specific graphics software to train machines on how to automatically recognize and classify objects with deep neural networks.

Images via Pinterest