Predicting Retention on the BBC Sounds Service Using Historic User Dynamics

In this 28-minute conference talk from DSC EUROPE 24, explore how BBC Sounds uses temporal machine learning models to predict user retention and reduce subscriber attrition. Learn about their approach to predicting the probability of users remaining subscribed across their 5 million user base. Discover how they combine dynamic temporal features (like minutes consumed) with static features (such as age and location) to forecast retention at 7, 30, and 90-day intervals. Compare the effectiveness of various temporal architectures including convolutional and recurrent neural networks against static models like XGBoost. Gain insights into optimizing temporal history usage for maximum model performance, which helps BBC Sounds tailor marketing efforts and reduce churn. This presentation by Ryan Timms was delivered on November 22nd at the Data Science Conference in Belgrade.