Optimizing Pinterest Shopping Recommendations for Purchases

Speaker's Bio:

Somnath Banerjee, Head of Shopping Discovery, Pinterest

Somnath Banerjee is the head of shopping discovery at Pinterest. He and the team are responsible for distributing personalized and inspiring shopping content across different surfaces of Pinterest. The team is spread across SF Bay Area and Toronto and is an expert in applying cutting-edge AI/ML techniques in content recommendation. Previously, Somnath was a director of machine learning and the head of search quality for Walmart.com. Somnath has decades of experience in working on machine learning and leading machine learning teams at various organizations. He holds a Ph.D. in Information Retrieval and has published a number of papers in prestigious conferences like SIGIR, WWW, etc. Somnath has been an invited speaker on E-commerce and machine learning at conferences worldwide; IDG at Seoul and AI Summit, Nvidia GTC, MLConf, etc.


Abstract:

Social platforms like Pinterest are becoming major discovery engines for online shopping. Everyday Pinterest recommends products and other shoppable content to millions of users on various surfaces e.g. Home Feed, Product Detail Page, etc. Different surfaces have different recommendation stacks optimized for different engagement objectives like "save", "click", “10sec video play” etc. Those objectives do not capture the primary goal of shopping to drive purchases. In this talk, I'll share how we include purchase optimization across different recommendation stacks of Pinterest. We optimize the item embeddings, generated using a transformer model, to create shopping candidates that are likely to drive purchases. We also develop a conversion model to predict the probability of checkout and use the predicted score for candidate ranking on different surfaces. Through this work, we are able to make changes in a few central places but able to increase purchases significantly from many different surfaces of Pinterest.