Re-examining assumptions in fair and unbiased learning to rank

In this thesis, we re-examine the assumptions of existing methods for bias correction and fairness optimization in ranking. Consequently, we propose methods that are more general than the existing ones, in the sense that they rely on less assumptions, or they are applicable in more situations. On the bias side, we first show that the click model assumption matters and propose cascade model-based inverse propensity scoring (IPS). Next, we prove that the unbiasedness of IPS relies on the assumption that the clicks do not suffer from trust bias. When trust bias exists, we extend IPS and propose the affine correction (AC) method and prove that, in contrast to IPS, it gives unbiased estimates of the relevance. Finally, we show that the unbiasedness proofs of IPS and AC are conditioned on an accurate estimation of the bias parameters, and propose a bias correction method that does not rely on relevance estimation. On the fairness side, we re-examine the implicit assumption that fair distribution of exposure leads to fair treatment by the users. We argue that fairness of exposure is necessary but not enough for a fair treatment and propose a correction method for this type of bias. Finally, we notice that the existing general post-processing framework for optimizing fairness of ranking metrics is based on the Plackett-Luce distribution, the optimization of which has room for improvement for queries with a small number of repeating sessions. To close this gap, we propose a new permutation distribution based on permutation graphs.