Inherent Trade-Offs in the Fair Determination of Risk Scores

One sense of fairness might be that the algorithm doesn’t take into account certain protected information, such as race or gender. Another sense of fairness might be that the algorithm is similarly accurate for different groups. Another notion of fairness might be that the algorithm is calibrated for all groups. In other words, it doesn’t overestimate or underestimate the risk for any group. Interestingly, any approach  that hopes to guarantee this property, might have to look at the protected information. So there are clearly some cases in which ensuring one notion of fairness might come at the expense of another.

In a recent paper, Professor Jon Kleinberg gave an impossibility theorem for fairness in determining risk scores. He shows that three intuitive notions of fairness are not reconcilable except in unrealistically constrained cases. So it might not be enough simply to demand that algorithms be fair. We may need to think critically about each problem and determine which notion of fairness is most relevant.



what does it mean for an algorithm to be fair? 


This article talks about how algorithms make decisions, how does machine learning relate to it, and it dissects many foundation ways algorithm and ML can go wrong and cause biases. It is a good indepth article on this topic.

University of Michigan - Ann Arbor

Bias is a systematic error in the estimate of an association between cause and effect. It may result from poor diagnosis/diagnostic criteria, poor case choice, poor choice of controls or variation in the way risk exposure is measured in case and controls.

Bias (Scientific)

http://approximatelycorrect.com/2016/11/07/the-foundations-of-algorithmic-bias/http://approximatelycorrect.com/2016/11/07/the-foundations-of-algorithmic-bias/

The Foundations of Algorithmic Bias


- Recall bias
- Selection bias
- Collider bias
- Omitted variable bias

Types of Bias

algorithm and bias

$\mathrm{bias}(\hat \theta_m) = \mathbb{E}(\hat \theta_m) - \theta$, 

where $\hat{\theta}_m$ is the the estimator of a paramter, $\theta$ is the true value of that parameter, and $m$ is the sample size/number of data points. 

The bias of an estimator for a parameter is the difference between the expected value that the estimator will take when trained on data and the actual value of the parameter used to generate the data.

Mathematical Defintion of Bias

A point estimator is a function that returns an estimate denoted by $\hat\theta$, using a given set of data as an input 

Learn Before

Related

Learn After