Learning Deep Probabilistic Embedding for Ad Creative Evaluation

Abstract

The performance and quality of ad creatives in online advertising can generally be evaluated by the rate at which the creatives are clicked on by users, and action rate such as the purchase of products are taken. On the other hand, such user behavior is highly noisy and uncertain, and it often depends on the potential attractiveness of the ad campaign and the target product of the ad. In this paper, we propose a new method for learning probabilistic ad creative embeddings to evaluate the ad creatives, which represent each ad creatives in the campaign as a Gaussian distribution in the latent space. Our probabilistic embedding can properly capture features of the creative from uncertain user behaviors and multiple ad creatives associated with the campaign. We evaluated our proposal using the record of real-world 200,000 ad creatives provided by Gunosy Inc. We confirmed that our probabilistic embeddings can accurately capture the serving performance of the ad creatives regardless of the text encoder (e.g., LSTM, BERT). Furthermore, we observed that our proposal achieves prediction with small uncertainty by using BERT, which is the recent state-of-the-art model.