With Spotify’s constant evolving product and fast growing user base, the user behavior in Spotify is becoming significantly diversified and complex. Thus, obtaining a holistic view of user behavior is crucial for future product development. However, such knowledge is difficult to obtain using A/B testing, the widely used technique to validate hypotheses, as A/B testing is designed to work well with relatively independent product changes and metrics. To address this issue, we recently adopted machine learning techniques to study the complex user behavior. The machine learning approach provides good understanding of relationships between various aspects of user behavior to the key metrics we are trying to optimize.
In this talk, Boxun Zhang gives an introduction of the Machine Learning approach, and some important lessons we have learned in the past, particularly about how Spotify combines the Machine Learning approach with traditional A/B testing.