Two Tencent IEG Papers Accepted at CIKM: Actor‑Critic Reinforcement Learning for Optimal Bidding and Adversarial Adaptation for Cross‑Domain Recommendation

Recently, Tencent IEG Growth Middle Platform team had two papers accepted at the top‑tier CIKM conference (CCF B class) in the fields of database, data mining, and information retrieval.

An Actor‑critic Reinforcement Learning Model for Optimal Bidding in Online Display Advertising

Abstract: In the real‑time bidding (RTB) scenario for online display ads, advertisers set bids for each impression opportunity. The paper formulates the bidding process as a constrained optimization problem and proposes an Actor‑Critic Reinforcement Learning (ACRL) model that incorporates user game‑behavior and media features, quantifies impressions, fits a Gaussian distribution of audience behavior, and has been deployed on Tencent’s gaming platform, influencing billions of daily impressions.

Cross‑domain Recommendation via Adversarial Adaptation

Abstract: Data sparsity and distribution shift hinder cross‑domain recommendation. The authors introduce an adversarial adaptation framework that aligns source and target domain distributions through a min‑max game between a domain discriminator and the target model, while calibrating target‑specific features. Experiments on large‑scale datasets show significant AUC improvements (e.g., +13.88% over PLE), and the method is now deployed on Tencent’s gaming platform.