Bias

Unfairness

Definition: The unjust or prejudicial treatment of different kinds of individuals from different categories/groups resulting in favor(benefits & opportunities) of a particular group. It is usually based on age, sex, skin color, language, economic condition, etc.

In artificial intelligence, it can be considered as an incorrect interpretation of the true relationship between the exposure and the outcome. Biases in the model can distort the true causal relationship.

A model is said to be biased if the predictor (y) is independent of the protected attribute (p) such that :

$$P(y | p) \neq P( y )$$

Importance:

1. Artificial Intelligence Models are increasingly being used in real-world use cases such as Loan Approval, Healthcare, Judiciary, etc which makes it imperative to on the AI community to minimize bias.

2. Learning more about bias helps us understand more about Human biases.

Where does bias come from?

1. Data Collection

2. Data inherently reflecting human bias(cognitive bias)

3. Biased Feedback loops

Removing Bias:

1. Awareness - find, understand and point out biases

2. Bias mitigation methods - Adding Fairness to the models, Datasheets for data sets & Model cards for Model reporting

3. Demonstrating Causal Relationships

Challenges:

1. More than 180 biases have been defined and classified and any one of which can affect the decisions we make.

2. Bias Feedback Loops