AI‑Driven Solutions for External Advertising Effectiveness at Alibaba Mama

Business background : With short‑video platforms such as Douyin and Kuaishou dominating user attention, advertisers on Alibaba Mama (Taobao & Tmall) seek to drive traffic from external media into the Alibaba ecosystem. The UD (UniDesk) external‑placement product enables merchants to allocate budgets across multiple off‑site media while providing comprehensive performance analytics.

Technical challenges : The core ocpx model requires accurate estimation of CTR (media side) and CVR (Alibaba side). Challenges include (1) missing front‑link user preference data due to privacy constraints, (2) extreme sparsity of post‑click conversion samples, (3) a mismatch between training (click‑only) and inference (full competition) spaces causing sample‑selection bias, and (4) the need for fine‑grained probability calibration for bidding.

Algorithmic solutions :

1) Federated learning (EFLS) : A privacy‑preserving framework where media partners share low‑dimensional embeddings of user preferences; Alibaba’s CVR model consumes these embeddings to improve ROI. The solution is open‑sourced on GitHub.

2) AutoHERI (CIKM 2021): An automated hierarchical representation integration model that aggregates features from upstream tasks to alleviate post‑click sample sparsity, using one‑shot architecture search.

3) UKD (WWW 2022): Uncertainty‑regularized knowledge distillation that introduces pseudo‑labels for unclicked samples, enabling full‑space CVR training while mitigating noise via uncertainty constraints.

4) AdaCalib (SIGIR 2022): A DNN‑based calibration module that learns feature‑wise calibration functions guided by posterior statistics, achieving tighter alignment between predicted probabilities and true conversion rates.

Future outlook : Ongoing work aims to unify multi‑scenario behavior‑sequence modeling and expand the external‑placement product suite, continuing the cycle of algorithmic innovation and business impact.