Problem Analysis and User Value Estimation in Advertising Scenarios

The presentation titled "Problem Analysis and User Value Estimation in Advertising Scenarios" is divided into three main parts: problem analysis of the placement scenario, practical user‑value modeling in advertising, and future work outlook.

It first examines the role of user value in growth, contrasting the classic AARRR acquisition model with the newer RARRA model that emphasizes activation and operation of existing users, and highlights the need for precise user‑value modeling as a foundational capability.

Key application scenarios include acquisition via CLTV modeling for channel‑level ROI estimation, cohort‑level CLTV analysis for bid strategy adjustment, and lifecycle operation where user‑value elasticity guides resource allocation.

The talk defines CLTV (Customer Lifetime Value) and CE (Customer Equity), explains how CLTV is modeled (SCV/nLTV), and describes challenges such as sparse samples, multi‑distribution data, and channel heterogeneity.

Evaluation methods for CLTV models are presented, including head‑% positive/negative sampling with AUC, Normalized Gini Coefficient, and the importance of ranking‑oriented metrics like Precise‑Recall for high‑value user selection.

Solutions to data sparsity and multi‑distribution problems are discussed, such as multi‑distribution LTV models (MDME), Bayesian sub‑component dependency models, and transfer learning with domain adaptation (DANN) to reuse data across channels.

Future work focuses on optimizing budget allocation based on customer‑asset models and continuously iterating user‑value models from sequence‑level to value‑level precision, leveraging advanced RTA capabilities for finer‑grained bidding strategies.

The content concludes with acknowledgments of the speaker, Xu Guoqiang, a senior researcher at Tencent, and the DataFun community.