As per the GDPR, the data controller must ensure a fair and lawful use of data. The analysis in [1] found three mean semantic notions of fairness in various translations of GDPR:

- correctness (Italian, Swedish, Romanian)
- loyalty (French, Spanish, Portuguese and the German version of “Treu und Glaube”) and
- equitability (French, Spanish and Portuguese).

All these three notions relates to the Roman law notion of “bona fide”. Thus the GDPR ensures that the data controller takes substantial mitigations of unfair imbalances that create situations of “vulnerability” for data subjects.

An investigation against Apple's credit card claims was opened by a US financial regulator (ref. dated November 11, 2019) due to complaints that the Apple's credit card offered different credit limits for men and women. It has been further followed by Apple's co-founder Steve Wozniak himself stating that the algorithms, ran by Goldman Sachs, set limits which might be inherently biased against women. Read full details [here].

The research conducted by M.I.T. Media Lab (ref. dated February 9, 2018), showed the examples of how bias has seeped into facial recognition softwares. E.g., Facial recognition algorithms made by Microsoft, IBM and Face++ were more likely to misidentify the gender of black women than white men. Gender was misidentified in up to 1% of lighter-skinned males in a set of 385 photos. Gender was misidentified in up to 7% of lighter-skinned females in a set of 296 photos ... read full story [here].

The article originally published by WIRE (ref. dated January 22, 2010) looked into several claims about how webcams and digital cameras had been unabashedly biased. Their face-tracking softwares refuses to track the faces of certain African-American and Asians. Read full story [here].

A study conducted by Georgia Institute of Technology (ref. dated March 6, 2019) stated that a dark skin person is more likely to be hit a by self-driving car than a white color person. This has raised questions in addition to worrying about how safe self-driving cars are, how they’d handle tricky moral trade-offs on the road, and how they might make traffic worse. Read full story [here].

Model Bias

Model bias is the presence of non-ethical discrimination in model predictions towards any of the data distribution groups. The data groups in which a bias can exits are commonly referred to as protected features, e.g., age, gender, ethnicity, as well as any proxy attributes such as postal codes as a representative of race.

Two of the civil rights organization namely, the Fair Housing Act (FHA) and Equal Credit Opportunity Act (ECOA) in USA have baselined race, skin color, gender, age, national origin, religion, familial status, disability, marital status, recipient of public assistance as protected features.

Bias Correction

Bias correction is a two step process in which a model is first scrutinized for bias against any protected features. We use more than 20 KPIs as discovered by state-of-art research, in order to detect model bias. The second step is to subject the model for bias correction which is a manual process and might result in the complete re-design of the original model.

Figure below shows an example of our approach towards correcting gender bias found in the churn & retention model of a moke-up study. Highest fairness score, or alternatively bias correction is achieved by taking different churn prediction thresholds for male and female customers.