Adaptive look-alike targeting in social networks advertising

摘要

Look-alike targeting models are based on an assumption that user similarity correlates with the probability of receiving the same positive feedback on advertising. They are successfully applied to searching and targeting of audiences in large-scale advertising networks. Usually, look-alike models lack an ability to improve their performance using experience gained during their application. Due to this, they highly depend on the initial user seed and can be thus affected by many negative factors including possible biases and noise presented in the seed. To minimize the influence of these factors, we suggest a look-alike model that is capable of adopting its targeting strategy to the feedback received. This model exploits Thompson Sampling algorithm applied to the space of users’ features. We have evaluated the proposed model in real advertising campaigns in a large social network assuming that its users were described by the online communities, members of which they were. Our method has achieved average 12.5% AUC improvement in comparison to the baseline look-alike models.