Sequence Optimization and Context-Aware CTR Re-Estimation for JD Advertising Ranking

The presentation introduces four main topics: (1) an overview of recommendation ad sorting techniques and technology choices, emphasizing the need to estimate traffic value and click‑through rate (CTR) accurately for better monetization.

(2) Context‑aware CTR re‑estimation, which uses a greedy forward search to treat each placed SKU as contextual information, allowing the model to re‑predict CTR with updated context.

(3) Reinforcement‑learning‑driven sequence optimization, describing two approaches—greedy forward search and global sequence evaluation via candidate set filtering—and a two‑step online rollout that first trains a sequence evaluation model and then iteratively improves a sequence generation model using Monte‑Carlo sampling.

(4) Session‑level ad auction mechanism optimization, proposing an incentive‑compatible, multi‑objective auction that integrates bid information with a learning‑based score, using a mixer‑MLP backbone to score each SKU per position and applying a second‑price style payment.

The Q&A section discusses how JD's rerank solution compares with industry peers, the use of virtual bids for natural traffic, multi‑metric reward design, and the observed revenue per mille (RPM) improvements from each iteration of the system.