Learning under Data Drift with Time-Varying Importance Weights