Understanding User Podcast Consumption Using Sequential Treatment Effect Estimation

Podcast recommendation systems are increasingly becoming prevalent with the growing popularity of podcasts. Accounting for the sequential nature of user interaction with such systems is cardinal to ensure that the recommendations are meaningful to the users. However, while there is growing work in effectively designing the recommendation systems, there is limited understanding of how the recommendations affect user behavior, i.e., what is the causal effect of a recommendation on long-term user behavior. In this work, we seek to explain the sequential causal effect of consuming a podcast category on long-term user behavior. We estimate the direct and indirect effect of user consumption, here, podcast category, on consecutive outcomes, i.e, aggregate weekly user podcast consumption observed across four weeks. We observe nonzero, non-homogeneous direct and indirect effects on outcomes across all weeks for multiple treatments, i.e. podcast categories. Moreover, consuming certain categories such as true crime have overall higher causal effects than others, highlighting that some podcast categories can alter user behavior resulting in long-term increased consumption.