Methods, computer systems, and computer storage media are provided for inferring sleep-related aspects for a user based, in part, on sensor data reflecting user activity detected by one or more sensors. In an embodiment, a user sleep model is trained using a dataset that includes previously-sensed data, descriptive information associated with the previously-sensed data, and/or interpretive data extracted from the previously-sensed data describing circumstances surrounding users when the data was acquired. In an embodiment, services providing time-sensitive recommendations personalized for a user's sleeping pattern using the inferred sleep-related aspects.