Pinterest is a mobile and web app that serves as a “catalog of ideas” (Pinterest CEO Ben Silbermann). Users submit images (with or without links and tags) to the app, and then collect each other’s images into their own lists of favorites. Its analog equivelant would be a cork board. A popular use is wedding planning, where a user currates a collection of images they are considering adopting for their own wedding.
Pinterest’s recommendation system—named Related Pins—is at the very heart of the business. As soon as a user logs on, they are presented with an infinite-scroll of images the recommendation system predicts the user will like. Users will also be recommended other users’ list of favorites that are similar to what the initial user has collected.
A good recommendation system is essential for the business to keep its users engaged and returning, and thus keeping ad money flowing to the company.
Several Pinterest data scientists have written a paper describing their experience and the recommendation system itself, which has been a very helpful resource for this essay.
Scenario Design
Who are your target users?
Young, social-media friendly people, especially those who have an ongoing creative hobby or poject, with disposable income. This project could be a wedding, but could also be fashion or cooking. A survey of actual Pinterest users shows they are disproportionately young, female, and from households with higher purchasing power. Two-thirds of pins ‘represent brands and products.’
What are their key goals?
Users want inspiration and help making their creative project a reality. They do this by collecting pins, and also through making purchases related to this pin. The same survey above notes that half of Pinterest users ‘have made a purchase after seeing a promoted [paid] pin’.
How can you help them accomplish their goals?
The best way to help them is to understand what their creative project is, along with their aesthetic-creative preferences, and showing them images consistent with those.
Reverse Engineer
Pinterest’s original Recommend Pins system was fairly simple. Each pin (or image/link), and its membership in users’ boards, is modeled a graph of pins and boards. The simplest way, then, to recommend other images is to examine the pins and boards that are often saved alongside the original pin.
User \(U_1\) might have a collection of hummus recipes. Another user \(U_2\) is browing hummus recipes, and saves one of them. Related Pins examines other pins that are often saved with it. It comes across \(U_1\)’s hummus recipes, notes that the recipe and these recipes are co-occuring within the same board, and recommends one of \(U_1\)’s hummus recipes to \(U_2\).
Pinterest has implemented other versions of Recommended Pins, though they are bit outside my scope of knowledge. The second verison used a ‘random walk service called Pixie,’ which ‘effectively computes Personalized PageRank on’ the graph. Another version uses something called Pin2Vec, which makes use of deep learning.
Improvements
It’s hard to recommend improvements to Pinterest—it does what it’s supposed to pretty well :). One thing I could think of is, when a user accesses one of their boards, they could specifically get recommendations for things within this specific board. However, I’m not sure they don’t do this already, and in fact it may be a superfluous functionality given all the other options Pinterest gives a user.