2019:KDD: Fairness-Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search

[Hamedloghmani]

Title: Fairness-Aware Ranking in Search & Recommendation Systems with application to LinkedIn Talent Search Year: 2019 Venue: KDD

Main Problem: Generally, when we are ranking people in real life, many factors might be the potential source of bias in the process of our decision such as looks, gender, behavior etc. The same issue is relatable in machine learning based decision making in ranking people and team formation. This problem can be tackled by three main approaches. First, by pre-processing which means the efforts before the training process such as dataset creation or feature engineering. Second, modifications dedicated to the model. And finally, post processing which includes calibrating the output of a potentially biased system in order to get unbiased results. This paper authors focused on the last approach.

Proposed Method: In order to make fairness definition more formal, they defined two conditions (original numbering is kept for ease):

r: A search request or a recommendation task A = {a1, . . ., al}: Set of disjoint protected attribute values (each candidate has exactly one value in A); Note that we denote the attribute value for candidate x as A (x), by abuse of notation. τr : Ranked list of candidates for r; τr [j] denotes jth candidate; τr k denotes the first k candidates in τr pq,r,ai : Desired proportion of candidates with attribute value ai that should be in the ranked list pτr ,r,ai : Proportion of candidates in τr with value ai count k (ai) denotes the number of candidates with attribute value ai among the top k results

The second condition is more important since it makes sure an attribute has at least the minimum representation that we expect. The authors came up with four different ranking algorithms. DetGreedy is an algorithm that first checks if any attribute is going to break the constraint #6, if so, it will choose the one with the highest score among them. Otherwise, the attribute with the highest next score from the list of attributes that have not satisfied their maximum requirement will be picked. DetCons and DetRelaxed: The authors propose these algorithms while pointing out that the previous one will easily fall into infeasible state. Their first step is just like the previous algorithm but if constraint #6 is not going to be violated, they favor one for the attributes that their minimum representation requirement is likely to be violated soon in the ranking. It is notable that “the above algorithms are feasible when the number of possible attribute values for protected attribute is less than 4.” DetConSort can be considered as a method to solve a general interval constrained sorting problem and the objective function is the maximization of the sorting quality. They argued that there is no single best algorithm and there is always a trade-off between utility and fairness. Dataset: Instead of using a dataset they used simulation because of two reasons. First, to be able to study settings where there could be several possible values for the protected attribute. Second, avoiding the challenge of position bias. In terms of fairness literature, the equalized odds metric is addressed in this work.

Input and Output:

Gaps: 1) Authors could not prove that the algorithms are feasible all the time. 2) The propossed methods are problem-specific. Taking the approaches that modifies the training process will be more generalizable. 3) They defined a fairness metric and measured fairness with it without proving how effective this measure would be in case of finding bias.

Codebase ( unofficial): https://github.com/yuanlonghao/reranking

[hosseinfani]

@Hamedloghmani thanks. please also add what the variables mean in (5) and (6) equations

[hosseinfani]

https://github.com/yuanlonghao/reranking/blob/main/examples/usage_example.ipynb

[hosseinfani]

@Hamedloghmani Based on the link you shared, https://developers.google.com/machine-learning/glossary/fairness, what type of fairness is addressed in this paper? equalized odds?