Cold-Start Optimization for Feed Ads: Algorithm Design and Experimental Evaluation

On November 29 (Sunday) from 09:30 to 10:30, DataFunTalk hosts a live broadcast featuring Dr. Zhang Renyu, an assistant professor at NYU Shanghai and economist/Tech Lead at Kuaishou.

Dr. Zhang’s research focuses on data‑driven optimization and A/B experimentation for large‑scale online platforms, with publications in top OR and MSOM journals and awards from INFORMS and POM. He teaches data science and operations research at NYU and Kuaishou, and has built economic and data‑science frameworks for evaluating and improving platform traffic and revenue, especially in recommendation and advertising systems.

The talk’s title is “Cold‑Start Optimization for Feed Ads: Algorithm Design and Experimental Evaluation.”

Cold‑start for feed ads (e.g., on Douyin, Kuaishou, Facebook) is a major challenge because limited data leads to low click‑through‑rate and conversion predictions for new ads, reducing bidding volume and potentially driving advertisers away. A data‑driven approach based on dual theory and neural bandits yields a new algorithm called Shadow Bidding with Learning (SBL). SBL adds a “shadow bid” on top of the advertiser’s original bid, enabling easy deployment in the ad system. Bilateral user‑ad experiments accurately measure the algorithm’s impact on short‑term spend and long‑term cold‑start value, solving the bias problem of one‑sided random experiments. Results show that SBL can significantly increase cold‑start value without sacrificing short‑term consumption.

Audience benefits:

All listeners gain a brand‑new algorithm design concept and experimental evaluation methodology.

Those unfamiliar with feed ads learn the business logic and cold‑start pain points of such ads.

New/Practical technical points introduced include cold‑start algorithms and their evaluation methods.