Beneath the hood of the Pinterest Associated Merchandise Engine

Pinterest gave a detailed look at the tech behind the associated product module, which displays Pinners' product ideas based on the Pins they are currently displaying. They are displayed as "Shop Similar" and "More to Shop" throughout the application.

The platform uses the current context, pin and user to generate a series of hundreds of products that might be relevant to that user at that point in time and then determines how those products are ranked in the most relevant way for each individual pinner.

Shopping Discovery Leader Somnath Banerjee, research intern at Pinterest Labs for Shopping Discovery Haoyu Chen, and software developer Pedro Silva wrote a detailed blog post on the process. Further highlights will follow.

You have divided related product recommendations into two high-level components: candidate generation and product ranking.

Candidate generation generates a set of hundreds of products that may be relevant to a particular pinner at that particular point in time based on the current context, PIN, and user.

Product ranking is carried out with a combination of multitask learning, calibration and Bayesian optimization to create a flexible, interpretable and scalable solution for candidate ranking.

Banerjee, Chen and Silva also explained the different types of engagement on the Pinterest platform: close up or tap on a pin to take a closer look; Storing or storing the pin on a board; Click or click the pin to visit the linked website. and long click when the user is away for a long time.

They wrote, “We used to treat predicting engagements as a binary classification task where impressions without engagements are negative and impressions with engagements are positive. We then selected an importance weight in the loss function for each job type according to their business values. For example, we can set a higher weight for long click engagement, as this indicates that the user may have made a purchase on the linked website. "

Banerjee, Chen and Silva explained the benefits Pinterest saw in moving to a multitasking model.

They wrote, “The advancement of the architecture of the related product ranking model from a single head model to a multitask neural network has improved engagement for all engagement types while providing us with more interpretable output that is useful not only for debugging purposes but also for model performance analysis. This work also allows us to further explore and expand the model to accommodate new tasks and new architectures as our use cases evolve. We also learned that offline Bayesian optimization doesn't always outperform other weight selection methods. However, the process gives us the assurance that the current weights are in line with our understanding of what pinners find appealing and inspiring in our retail space. We will continue this work using an online Bayesian optimization approach that has proven far more successful than its offline counterpart in solving similar problems. "

Pinterest also shared some dates related to the holiday shopping season:

  • Starting in September, the number of pinners engaged in retail space increased by 85% over a six-month period.
  • The holidays were in the foreground earlier in the year than usual. Search for Christmas gift ideas rose three times as much in April as it did in the same month last year, while searches for public holidays in general increased by 77%.
  • Search for personalized Christmas gift ideas is up 46% year over year.
  • The search for luxury gifts has increased three times year-on-year.
  • Pinterest has also seen "significant growth" in finding popular shopping categories such as beauty, home decor and womenswear.

Comments are closed.