Post-Estimation Smoothing: A Simple Baseline for Learning with Side Information