Modeling delayed feedback in display advertising

In performance display advertising a key metric of a campaign effectiveness is its conversion rate -- the proportion of users who take a predefined action on the advertiser website, such as a purchase. Predicting this conversion rate is thus essential for estimating the value of an impression and can be achieved via machine learning. One difficulty however is that the conversions can take place long after the impression -- up to a month -- and this delayed feedback hinders the conversion modeling. We tackle this issue by introducing an additional model that captures the conversion delay. Intuitively, this probabilistic model helps determining whether a user that has not converted should be treated as a negative sample -- when the elapsed time is larger than the predicted delay -- or should be discarded from the training set -- when it is too early to tell. We provide experimental results on real traffic logs that demonstrate the effectiveness of the proposed model.