Modeling Advertising Attractiveness: Data Analysis, Pairwise Learning, and DeepFM Optimization

High‑quality video content can boost user engagement on platforms such as WeChat Channels and Douyin, while effective ad videos increase watch time and conversion. Tencent Ads has built an in‑house attractiveness estimation model to predict how compelling an ad video is, thereby improving ad quality and user experience.

The author, a senior algorithm researcher at Tencent Ads Multimedia AI Center, explores both data‑driven and modeling approaches for the attractiveness dimension of ad quality. The focus is on the 3‑second completion rate (3s‑play), which correlates strongly (r=0.7) with click‑through rate and serves as a proxy for user attraction.

Background : Advertisers often test many creative assets, incurring 10‑20% trial costs. Ranking new assets by predicted quality can reduce these costs. Existing works such as Facebook Creative Compass and ByteDance PEAC evaluate creative quality using various metrics.

Related Work : Summarizes prior creative quality metrics from Facebook, ByteDance, Baidu, and Google, highlighting differences between video and search ad contexts.

3.1 V0 – Pairwise MLP Model : To avoid costly manual labeling, the authors use 3s‑play as an implicit label. A pairwise MLP predicts which of two creatives will achieve a higher 3s‑play, achieving 60.8% accuracy on WeChat Moments ads.

3.2 V1 – Pairwise DeepFM Model : Incorporates additional factors such as exposure time and audience demographics, which were shown to affect 3s‑play. DeepFM handles sparse features and improves accuracy, especially when exposure‑time features are added (+6.1%).

3.3 V2 – Hierarchical Sampling : Extends pairwise training beyond the same ad by constructing sample pairs at three levels – within the same ad, within the same account, and within the same traffic slot. This hierarchical sampling raises accuracy from 74.0%/73.0% to 76.9%/89.0% on account and traffic test sets.

3.4 Model Interpretability : Uses the final MLP layer for clustering; visualizes clusters with differing average 3s‑play and predicted scores. To reduce visual inconsistency, a consistency loss aligns visual semantic similarity with clustering similarity, improving intra‑cluster visual coherence.

4. Business Deployment : The attractiveness model is integrated into Tencent’s ad system, enhancing material recommendation (higher CTR and 3s‑play) and ad ranking (improved GMV). It runs on the TaiChi ML platform with the Hunyuan AI large model as backbone.

5. Summary & Outlook : The model reduces labeling cost, captures diverse influencing factors, achieves up to 97.5% accuracy on traffic‑level comparisons, and opens avenues for multi‑task quality modeling, combined material‑audience scoring, and deeper interpretability.

References : Includes citations to Facebook Creative Compass, ByteDance PEAC, Zhao et al. (2019), and Guo et al. (DeepFM, 2017).