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

Guaranteed delivery (GD) and real-time bidding (RTB) constitute two parallel profit streams for the publisher. The diverse advertiser demands (brand or instant effect) result in different selling (in bulk or via auction) and pricing (fixed unit price or various bids) patterns, which naturally raises the fusion allocation issue of breaking the two markets' barrier and selling out at the global highest price boosting the total revenue. The fusion process complicates the competition between GD and RTB, and GD contracts with overlapping targeting. The non-stationary user traffic and bid landscape further worsen the situation, making the assignment unsupervised and hard to evaluate. Thus, a static policy or coarse-grained modeling from existing work is inferior to facing the above challenges.

This paper proposes CONFLUX, a fusion framework located at the confluence of the parallel GD and RTB markets. CONFLUX functions in a cascaded process: a paradigm is first forged via linear programming to supervise CONFLUX's training, then a cumbersome network distills such paradigm by precisely modeling the competition at a request level and further transfers the generalization ability to a lightweight student via knowledge distillation. Finally, fine-tuning is periodically executed at the online stage to remedy the student's degradation, and a temporal distillation loss between the current and the previous model serves as a regularizer to prevent over-fitting. The procedure is analogous to a cascade distillation and hence its name. CONFLUX has been deployed on the Tencent advertising system for over six months through extensive experiments. Online A/B tests present a lift of 3.29%, 1.77%, and 3.63% of ad income, overall click-through rate, and cost-per-mille, respectively, which jointly contribute a revenue increase by hundreds of thousands RMB per day. Our code is publicly available at https://github.com/zslomo/CONFLUX.