How Pinterest uses AI for your recommendations
Apps like Pinterest, Instagram, Netflix, and even news sites personalize your landing page to match your preferences. How do they decide what to show you? UAntwerp’s newest honorary doctor Jure Leskovec explains how AI models can provide these tailored recommendations.
How AI Improves the User Experience on Pinterest
Computer scientist Jure Leskovec is a professor at Stanford University, an honorary doctor at UAntwerp, and a leading researcher in industry. As Chief Scientist at Pinterest, he studied how artificial intelligence can enhance user experience. “Pinterest is an online platform where people come to be inspired, often for creative projects. They see images that they can save or ‘pin’ to their own ‘boards’: visual collections built around certain themes, like home decor.”
Leskovec explains that every image a user sees is the result of an AI-driven recommendation. Together with his team, he improved the AI model that decides what to show to which users. “The old model considered each image and its features individually to determine whether to show it to you. We optimized this by treating Pinterest as one huge network of pictures that are all connected based on how often users save them together on boards.” This network representation captures user behavior in a way that improves the AI model: “While computers don’t understand pictures like we do, they can learn from our behavior. For instance, rugs and tapestries may look similar, but people rarely save them on the same board. So, most users don’t want a tapestry recommended after saving pictures of rugs. The way people group images helps the model understand them in a more nuanced way and provide better recommendations.”
How Pinterest Prevents Clickbait
Another key aspect of generating effective recommendations lies in clearly defining your purpose, Leskovec says. “We can create AI models that predict users’ next actions after seeing a particular item: clicking on it, saving it, or – for example in web shops – adding it to their shopping cart. It’s essential to know which of these actions you want to encourage.” That choice depends on the platform’s goal and philosophy. “Pinterest, for instance, does not aim for maximum clicks. That would lead to feeds full of clickbait: images that grab your attention but aren’t what you were really looking for. For Pinterest, user satisfaction means people saving images to their boards. That principle impacts which items the AI model recommends.”
Overprotecting your work only slows down progress. Open science drives innovation!
From Pinterest to Facebook: The Power of Collaboration
Leskovec’s network-based recommendation technique is now used by several major companies where accurate suggestions are essential – including Facebook, Uber, and Amazon. It’s no surprise that Leskovec advocates for closer ties between academia and industry, believing that such partnerships keep researchers focused on real-world problems and benefit both sides. He also greatly values open science. “Computer science is a very transparent field that embraces an open-source culture, even within industry. At Pinterest, we published papers on our advancements and made the programming code and part of the dataset freely available. Overprotecting your work only slows down progress. Open science drives innovation!”