查看文章

In online advertising, click-through rate (CTR) prediction typically utilizes click data to train models for estimating the probability of a user clicking on an item. However, the different presentations of an item, including its position and contextual items, etc., will affect the user's attention and lead to different click propensities, thus the presentation bias arises. Most previous works generally consider position bias and pay less attention to overall presentation bias including context. Simultaneously, since the final presentation list is unreachable during online inference, the bias independence assumption is adopted so that the debiased relevance can be directly used for ranking. But this assumption is difficult to hold because the click propensity to the item presentation varies with user intent. Therefore, predicted CTR with personalized click propensity rather than debiased relevance should be closer to real CTR. In this work …