DrEM framework teaches video recommenders to handle noisy preference predictions

Industrial video recommendation systems rely on a multi-stage pipeline: an upstream multi-task model outputs multi-dimensional user preference predictions (pxtrs), which an ensemble ranking stage fuses into a single score. The problem is that pxtrs carry prediction noise, and because downstream models treat them as ground truth, the noise propagates on two sides. On the supervision side, noisy pxtrs can flip proxy preference labels and inject erroneous gradients; on the feature side, noise destabilizes the ranking score itself. Existing methods overlook this, but DrEM, proposed by Huang et al., directly attacks both failure modes.

DrEM introduces a risk-denoising robust loss that corrects empirical risk using an estimated preference-flip probability, effectively cleaning supervision signals. For feature robustness, it samples perturbations from the predicted noise distribution and applies a preference-preserving ranking consistency regularizer, maintaining stable outputs. The authors also derive an approximate noise distribution and prove the robust loss remains effective even when flip probability estimation is imperfect. Combined with offline experiments and large-scale online A/B tests, DrEM demonstrates measurable gains in ranking quality, making it a practical upgrade for real-world video recommendation systems.