Abstract: Signals representing local events and/or state are captured at a mobile device and utilized by a machine learning system to recognize patterns of user behaviors and make predictions to automatically launch an application, initiate within-application activities, or perform other actions. The local signals may include, for example, location information such as geofence crossings; alarm settings; use of network connections like Wi-Fi, cellular, and Bluetooth®; device state such as battery level, charging status, and lock screen state; device movement indicating that the device user may be driving, walking, running, or stationary; audio routing such as headphones being used; telemetry data from other devices; and application state including launches and within-application activities. A feedback loop is supported in which the machine learning system may utilize feedback from the user as part of a learning process to adapt and tune the system's predictions to improve the relevance of the predictions.