Advertising Targeting: Pain Points, Audience Insight System, Relationship Network Applications, and Modeling Strategies

The presentation begins by outlining the main pain points in advertising: traffic concentration among a few large platforms, escalating competition and costs, and data black holes that limit effective audience profiling.

To address these issues, a three‑layer audience insight system is introduced. The bottom layer aggregates data from media platforms, internal business sources, and external providers. The middle layer hosts commonly used advertising models, and the top layer generates audience packages for targeting and exclusion.

The discussion then shifts to relationship network applications, describing both supervised (hand‑crafted features, random‑walk/graph‑embedding, end‑to‑end GCN) and unsupervised (graph clustering) methods, and illustrating how these techniques can improve user segmentation and risk control.

Subsequently, various modeling strategies are compared: single‑stage models for each funnel step, cross‑stage models that span multiple steps, and multi‑task learning approaches such as MMOE and ESSM. The benefits of media‑joint modeling platforms for rapid model development and the challenges of feature and algorithm limitations are also highlighted.

Finally, the talk covers attribution challenges across multiple touchpoints, the impact of team structure on modeling, and emphasizes the importance of data quality, preprocessing, and collaboration with data providers to enhance model performance.