CONFLUX: A Request-level Fusion Framework for Impression Allocation via Cascade Distillation

The paper introduces CONFLUX, a request-level fusion framework for allocating impressions between contract and real-time bidding (RTB) ads, presented at KDD 2022.

It addresses three main challenges: unsupervised training due to lack of per-impression labels, strict service latency requirements, and model degradation from distribution shift.

The solution uses linear programming to generate supervised training samples, then trains two sub-networks (RTB-net and GD-net) to predict winning probabilities, and applies cascade distillation (Teacher-net to Student-net) with logits, hints, and similarity losses to reduce model size while preserving accuracy.

Online calibration employs temporal distillation loss to adapt to drifting data without forgetting.

Offline A/B tests on splash, pre-roll, and in‑feed ad logs show significant revenue gains; online deployment for over six months yielded a 3.29% increase in contract ad consumption, 1.77% rise in click‑through rate, and 3.63% higher CPM.

The authors conclude that CONFLUX achieves a win‑win for advertisers and the platform by balancing impression allocation and maintaining ad effectiveness.