Full‑Chain Linkage Techniques for Alibaba Mama Display Advertising: From Precise Value Estimation to Set‑Selection Models

Introduction – The deep‑learning era initially brought huge technical and compute benefits to display advertising, but as these advantages fade, individual modules in the cascade ranking system (recall, pre‑ranking, ranking, re‑ranking) encounter performance bottlenecks.

Problem and Challenges – Gaps between upstream modules and downstream objectives, as well as sample‑selection bias (SSB), limit further gains. The author proposes a full‑chain linkage strategy to align objectives across the entire pipeline.

Technical Solutions Overview

1. Precise‑Value Estimation – The Point‑based Deep Match (PDM) model enables full‑library vector recall for arbitrary targets (e.g., RPM, GMV) by converting eCPM into inner‑product form using ALSH. The Entire‑Space Domain Adaptation Model (ESDM) addresses pre‑ranking SSB by joint training with the real‑time deep fully‑connected model (COLD) and knowledge distillation from the ranking model.

2. Set‑Selection Techniques – Learning‑to‑Rank based Deep Match (LDM) for recall learns the final ranking order as the target, mitigating bid‑sensitivity issues. The Learning‑to‑Rank based and Bid‑Sensitive Deep Pre‑Ranking Model (LBDM) extends this to pre‑ranking, introducing a bid‑monotonic pairwise loss to preserve bid‑sensitivity while aligning with downstream objectives.

Exploit & Explore Full‑Chain Channels – The system separates short‑term exploitation (set‑selection focus) from long‑term exploration (precise‑value estimation focus), allowing independent optimization of immediate performance and ecosystem health.

Results – Online experiments show consistent improvements: PDM (CTR +1.5%, RPM +2%), ESDM (CTR +3%, RPM +1.5%), LDM (CTR +3%, RPM +4%), LBDM (CTR +8%, RPM +5%). The full‑chain linkage has been deployed across all core information‑flow ad scenarios, delivering over 10% revenue growth.

Summary & Outlook – By combining precise‑value estimation and set‑selection routes, the full‑chain linkage aligns the entire cascade with final system goals, reduces loss, and opens new growth space. Future work includes deeper module fusion, end‑to‑end ranking, and tighter integration of the two technical routes.