Event‑Aware Graph Extraction and Adaptive Clustering‑Gain Network for Insurance Creative Recommendation

Background Insurance creative recommendation is a vertical application of advertising creative, facing large‑scale optimization under complex constraints such as sparse user and creative data, counterfactual exposure, and cross‑industry cold‑start.

Challenges The work identifies three major challenges: (1) data sparsity on both user and creative sides, (2) counterfactual sample problem where each user only sees one creative, and (3) cold‑start across different insurance industries.

Solution Overview To tackle these issues, the authors propose three ideas: (a) group‑level insights to alleviate individual sparsity, (b) random exposure of new creatives to collect feedback, and (c) leveraging event‑level similarities across industries via a heterogeneous graph and uplift learning.

Network Architecture The proposed system consists of three parts: (1) Event‑aware graph vector extraction – offline pre‑training to embed user, context, and creative features; (2) Adaptive clustering‑gain network – uses uplift thinking to aggregate group preferences for individual prediction; (3) Co‑Attention between user, event, and creative embeddings to capture pairwise interactions.

Heterogeneous Graph Construction (EAGT) Nodes represent users, events, and creatives; edges encode conditional conversion probabilities (e.g., higher weight for a user node under adverse weather). Node embeddings are learned via self‑supervised edge prediction, and domain‑shared embeddings enable knowledge transfer across similar industries.

Training Losses Four loss functions are designed: intra‑loss for clustering balance, cross‑entropy loss for the clustering‑gain network, a global cross‑entropy loss, and a fused loss that combines all three.

Experiments and Analysis The model is evaluated on a mixed industrial and public dataset (Alibaba Feizhu and Tianchi advertising creative data). Compared with baseline ranking algorithms and other industry methods, the proposed architecture improves AUC, with the multi‑view attention network contributing the most, followed by the gain network and the heterogeneous graph.

Conclusion and Outlook The work introduces two innovations: an event‑aware graph extractor that incorporates cross‑scene event information into creative recommendation, and an adaptive clustering‑gain network that mitigates counterfactual bias using group wisdom. Future directions include further refinement of event modeling and broader cross‑industry transfer.