The Difference Between a Click and a Cart-Add: Learning Interaction-Specific Embeddings

For large-scale online marketplaces with over millions of items, users come to rely on personalized recommendations to find relevant items from their massive inventory. One hallmark of the shopping experience in such online marketplaces is the many ways a user can interact with an item they are interested in: they can click it, favorite it, add it to cart, purchase it, etc. We hypothesize that the different ways in which a user interacts with an item indicates different kinds of intent. Consequently, a user’s recommendations should be based not only on items from their past activity, but also the way in which they interacted with these items. Co-occurrence based methods have been successfully used to give recommendations that incorporate interaction types, such as the popular “Because you purchased X, you may also purchase Y”. In this paper, we propose a novel method that generalizes upon the co-occurrence methods to learn interaction-based item embeddings that encode the co-occurrence patterns of not only the item itself, but also the interaction type. The learned embeddings provide a convenient way of approximating the likelihood that one item-interaction pair would co-occur with another by way of a simple inner product. To show their effectiveness, we deploy the interaction-based embeddings in an industry-scale recommendation system that serves live traffic on Etsy.com. We find that taking interaction types into account shows significant improvements in accurately modeling user shopping behavior for both online and offline metrics.