Duolingo uses a machine learning algorithm to decide which push notification to send each user, at what time, and how often — optimizing for the messages most likely to get a lapsed learner back into the app. The system was published as research in 2020 and was shown to increase the number of daily active users by 0.5% and improve retention of new users by 2%.

Published at the ACM SIGKDD 2020 data science conference, Duolingo's "Sleeping Recovering Bandit" algorithm processes approximately 200 million notification examples to learn optimal notification strategies per user. A key technical challenge was accounting for "novelty effects" — users becoming desensitized to repetitive messages over time — which the algorithm handles through recovery periods. The system also manages the conditional eligibility problem: not every notification is appropriate for every user at every time (e.g., a user mid-lesson should not receive a reminder). Different notification messages were found to perform differently across languages and user segments. Duolingo released the dataset of 200 million notification examples publicly.