Advertising Creative Optimization Using Hybrid Bandit Models

This content discusses the optimization of advertising creatives using hybrid bandit models, focusing on how machine intelligence can enhance the selection and effectiveness of advertising creatives.

Advertising creatives are defined as "ideas" or "points" that advertisers use to attract consumer attention by presenting product features in an appealing way. These creatives can include images, titles, videos, and live streams.

The Alibaba Moments Advertising Creative Optimization Platform uses algorithmic capabilities to help advertisers efficiently and accurately select the most attractive creatives. This involves two main aspects: optimizing complete creatives and dynamically optimizing creative elements.

For complete creative optimization, the problem is modeled as an Exploration & Exploitation (E&E) problem, where each candidate creative is given a chance to be displayed to users and receive feedback. The system then allocates more traffic to the creatives that have shown the best performance so far. This E&E strategy is solved using bandit models like Smoothed-Greedy, Thompson Sampling, and LinUCB.

However, standard bandit strategies face challenges such as cold starts, where data is insufficient for initial decisions. To address this, a hybrid bandit model with visual priors is proposed. This model combines a Visual-aware Ranking Model (VAM) and a Hybrid Bandit Model (HBM). The VAM learns visual features correlated with click-through rates from sufficiently deployed data, while the HBM uses these features and model parameters as priors, updating them based on actual deployment data.

The content also covers dynamic creative optimization, where the system receives creative elements instead of complete creatives. These elements are then optimized based on user behavior and deployment results to balance visual and deployment effects.

Experiments are conducted on datasets like Mushroom and CreativeRanking to validate the effectiveness of the proposed models. The results show that the hybrid model outperforms other methods in terms of regret and click-through rates, especially in cold start scenarios and long-term performance.