Under the Hood of Pinterest’s Related Products Engine

Pinterest provided a detailed look at the engineering behind its related products module, which shows Pinners product ideas based on the Pins they are viewing at the time, appearing across its application as “Shop Similar” and “More to Shop.” The platform uses the current context, Pin and user to generate a set of hundreds of products that might be relevant to that user at that moment, and it then determines how to rank those products in the most relevant way for each individual Pinner. Head of shopping discovery Somnath Banerjee, Pinterest Labs research intern for shopping discovery Haoyu Chen and software engineer, shopping discovery Pedro Silva penned a detailed blog post on the process, and highlights follow. They broke down related products recommendations into two high-level components: candidate generation and product ranking. Candidate generation is the process of generating a set of hundreds of products that may be relevant to a specific Pinner at that particular moment, based on current context, Pin and user. Product ranking is done with a combination of multitask learning, calibration and Bayesian optimization to build a flexible, interpretable and scalable candidate ranking solution. Banerjee, Chen and Silva also detailed the different types of engagement on the Pinterest platform: close-up, or tapping on a Pin to take a closer look; save, or saving the Pin to a board; click, or clicking through the Pin to visit the linked website; and long-click, when the user stays offsite for an extended period of time. They wrote, “Previously, we treated engagement prediction as a binary classification task where impressions without engagements are negatives, and impressions with any engagement are positives. We then chose an importance weight in the loss function for each engagement type according to their business values. For example, we may choose to set a higher weight for the long click engagement because it indicates that the user might have made a purchase on the linked website.” Banerjee, Chen and Silva explained the benefits Pinterest saw when it moved to a multitask model. They wrote, “Evolving the related products ranking model’s architecture from a single head model to a multitask neural network has improved engagement across all the engagement types, while also providing us with more interpretable outputs, which are not only useful for debugging purposes, but also for model performance analysis. This work also allows us to continue exploring and expanding the model to accommodate new tasks and new architectures as our use cases evolve. We also learned that offline Bayesian optimization does not always outperform other weight selection methods. However, the process gives us confidence that the current weights are aligned with our understanding of what Pinners find engaging and inspirational on our shopping surfaces. We’ll follow up this work with an online Bayesian optimization approach, which has proven far more successful for solving similar problems than its offline counterpart.” Pinterest also shared some data related to the holiday shopping season: As of September, the number of Pinners engaging with shopping surfaces was up 85% over a six-month period.The holidays were top of mind earlier in the year than usual, with searches for Christmas gift ideas up three times in April compared with the same month in previous years, while holiday searches in general were up 77%.Searches for personalized Christmas gift ideas are up 46% year-over-year.Searches for luxury gifts are up three times year-over-year.Pinterest has also seen “significant growth” in searches for popular shopping categories such as beauty, home décor and women’s fashion.