CTR-Driven Advertising Text Generation and Bundle Creative Optimization (CREATER & CONNA)

This article presents two research works—CREATER and CONNA—developed by Alibaba’s advertising team for dynamic creative optimization in external advertising platforms.

Background : Creative elements directly affect ad efficiency and user experience. The team distinguishes between Programmatic Creative (mass production of diverse creatives) and Dynamic Creative Optimization (DCO), which iteratively refines creative components (titles, copy, images, videos) using online feedback.

CREATER Model : A CTR‑driven text generation model for low‑resource scenarios. It uses user reviews as source text, aspect terms as control codes, and incorporates CTR signals via contrastive fine‑tuning. Training consists of two stages: (1) Controlled Pre‑training on massive review data with a custom self‑supervised objective that masks aspect‑related spans and forces the model to reconstruct them; (2) Contrastive Fine‑tuning that leverages paired A/B test copies (positive/negative CTR) to encourage high‑CTR generation. The architecture is a standard Transformer encoder‑decoder with control codes prepended to the masked source.

Experimental Results (CREATER) : Offline metrics (ROUGE, BLEU) and online A/B tests show significant CTR and CPC improvements over baselines (PGNet, Transformer, RL‑based methods). Ablation studies confirm the benefit of aspect‑controlled masking and contrastive loss.

CONNA Model : A non‑autoregressive bundle creative generation model that jointly predicts a set of heterogeneous elements (products, slogans, templates). It encodes user history, recent clicked items, and candidate element sets with type embeddings, then decodes all positions in parallel. To handle the unordered nature of element sets, a set‑based loss with Hungarian matching is used, avoiding penalization for correct but permuted outputs. Contrastive loss further distinguishes high‑CTR (positive) from low‑CTR (negative) bundles.

Experimental Results (CONNA) : Offline evaluations on large and small training splits demonstrate superior creative quality, element diversity, and decoding speed compared to autoregressive baselines (MultiDec, UnifiedDec, PointerNet, RL‑Transformer). Online A/B tests on major sales events report CTR gains of up to +8% and CPC reductions of up to -13%.

Conclusion & Future Work : Both CREATER and CONNA are simple yet effective solutions for DCO, achieving strong online performance. Future directions include incorporating contextual factors (weather, location, time) and enriching creative elements with fine‑grained tags for better personalization.