Categorizing User Sessions at Pinterest

Abstract

There are many different ways for individuals to use a given Internet application. By recording and analyzing detailed logs of user in-application activity, we identify patterns of behavior that reflect distinct use cases. Not only does this lead to a better understanding of user intent, but also generates important insights with actionable product and business outcomes. In this application oriented paper, we study the problem of user session classification, where the goal is to discover categories of user in-application behavior using event logs, and then consistently label user sessions on an ongoing basis. We develop a practical three-step approach which uses clustering to discover categories of sessions, builds classifiers to label new sessions, and finally performs daily classification in a distributed pipeline. An important innovation of our approach is replacing a set of events identified as long-tail features with a new feature that is less sensitive to product experimentation and logging changes. This allows for robust and stable identification of session types even though the underlying application is constantly changing. We apply the approach to Pinterest sessions and verify the results with a user survey. Our solution classifies millions of user sessions daily and leads to actionable insights that guide business decisions relating to product design, monetization, and growth.