We tried some ideas inspired by this paper https://arxiv.org/abs/1608.04468 . We also tried other schemes trying to debias the training data based on impression priors. So far they have not lead to any improvements. This may be an issue that is very product dependent. For products where the searcher is looking for information and the product is free to use, positional bias may be high as users just click the top few results. In a product like Airbnb where transactions run into hundreds of dollars, users are more engaged and check through the results more carefully. But the question is far from settled and we are still trying to understand this better.