Hybrid Bandit and Visual-aware Ranking Models for Advertising Creative Selection and Dynamic Optimization

Advertising creative selection aims to choose the most attractive combination of images, titles, videos, and other assets for a given product. Traditional manual selection is limited, while machine intelligence can learn from massive data to rank and optimize creatives.

The authors model complete‑creative selection as a standard Explore‑and‑Exploit (E&E) problem using bandit strategies (e.g., Smoothed‑Greedy, Thompson Sampling, LinUCB) and introduce visual priors to improve cold‑start performance. They propose a Visual‑aware Ranking Model (VAM) that extracts high‑level visual features via deep convolutional networks and learns a scoring function using point‑wise and list‑wise losses.

To overcome VAM’s deterministic exploitation and global‑pattern bias, a Hybrid Bandit Model (HBM) is built on top of the visual features. HBM treats the click outcome as a linear function of learned visual embeddings with Bayesian linear regression, enabling posterior sampling for balanced exploration and exploitation.

Beyond whole‑creative ranking, the paper describes dynamic creative optimization, where individual creative elements (templates, images, titles, benefit points, call‑to‑action) are personalized per user. A multi‑scale DNN predicts CTR from user, ad, and element features, and a tree‑structured representation of element‑parameter combinations allows efficient optimal path search via dynamic programming.

Feature interaction is further enhanced by automatically searching for the best interaction function among concatenation, element‑wise multiplication, addition, max, and min, using a one‑shot NAS‑style algorithm to assign importance weights to each candidate.

Extensive experiments on two large internal datasets (50 M ads, 1.7 M creatives) and public benchmarks demonstrate that the VAM‑HBM pipeline achieves lower regret and higher sCTR than baselines, improves cold‑start CTR by 5 %, and that the AutoCO interaction search yields up to 7 % online CTR lift.

The work concludes with references to related conference papers and notes the broader AI‑driven advertising products developed by the Alibaba‑Mama creative algorithm team.