In one embodiment, a method includes receiving first data indicating a current usage of a mobile device by a user. The method includes accessing second data associated with past user states of the user, where each user state corresponds to a temporal, spatial, or modal accessibility of the user. A future user state of the user is determined based on a weighted set of predictor functions, each predictor function comprising a machine-learned algorithm for a particular user state. The probability of each user state is based on past user states of the second data, wherein each past user state is weighted by a decay factor based on the corresponding time. One of the user states is selected as the future user state, and the operation of the mobile device is adapted at a future time to account for that future user state.