<aside>

Summary

Any solution of bias starts with awareness of the bias’ existence

</aside>

<aside>

Readings

Extra article about the two types of worldviews

</aside>

<aside>

Keywords -

• Bias - Focuses on representation (of culture, gender, race)

• Fairness - Focuses on the decision outcome

</aside>

<aside> 💡

Two types of Worldviews

What you see is what you get

facts about my character SHOULD BE CONSIDERED when making a decision. eg. bank loan: if there are two identical people with the same risk factor, etc then they should both be approved/denied similarly to be considered fair

• WYSIWYG guarantees individual fairness

When you train a model on data, you might be making the implicit assumption that data is representative of the qualities you are interested in modeling.

We are all equal (Structural Bias)

The whole idea of equal opportunity and other group fairness metrics is to explicitly encode that different groups should have equal treatment since any differences observed are the result of factors outside of an individual’s control

• Attempts to achieve individual fairness leads to discrimination </aside>

<aside> 💡

Categories of Fairness

Individual

Group

Subgroup

</aside>

<aside>

Metrics of Fairness

• Equal Odds
• Equal opportunity
• Demographic parity
• Fairness through awareness
• (Group Unaware)Fairness through unawareness: Removing defining information from dataset eg. gender data
• Treatment equality
• Test fairness </aside>

<aside>

Confusion Matrix

</aside>

Example usage

<aside> 💡

Fairlearn (by Microsoft)

</aside>

<aside> 💡

The impossibility of satisfying more than one type of fairness

</aside>