Algorithmic Bias Mitigation

Definition

Reducing bias doesn't start with the model, but with understanding the problem and the data. We need to examine which groups are underrepresented, which attributes may lead to discrimination, how errors are distributed, and which fairness criteria are important. Data cleaning, balancing, restrictions, auditing and monitoring are then applied.

Beispiel

In the hiring model, the team checks whether the quality of recommendations for candidates from different regions, age groups or educational trajectories is degraded.

Warum es wichtig ist

The term is important for responsible AI: a model may be accurate on average, but unfair for certain groups.

So funktioniert es

Approaches can be before training, during training and after it: improving data, changing the loss function, calibrating results and regular auditing.

Wo es genutzt wird

• recruitment and HR
• credit scoring
• moderation and recommendations

Einschränkungen

You cannot simply “remove a sensitive sign” and consider the problem solved: other signs can indirectly reproduce it.

FAQ