Delayed Feedback with Selection Bias in Recommender Systems

One challenge for online advertising companies is the lack of time-sensitive data, as users may not buy the product immediately after clicking on the ads. Wanting to improve the performance of their products, advertisements companies can constantly train on the real-time data stream. Without fully generated data, this approach will lead to few biases. In this paper, we discussed non-parametric models regarding delay distributions and three estimators for the training model. In the end, we proposed the Dual Learning Algorithm. While training on 100000 data points and different sets of hyper parameters, we achieved the overall best set of F1 and AUC scores: 0.55 and 0.70.