Fairness/Bias Monitoring

Fairness or bias can be broadly defined as the absence of discrimination or favouritism toward a person or group based on their characteristics. Even with perfect data, our modelling techniques may still result in bias. In its simplest terms, bias is the situation where the model consistently predicts distorted results because of incorrect assumptions. When we train our model on a training set and evaluate it on a training set, a biased model produces significant losses or errors.

Data scientists and ML engineers can regularly check predictions for bias with the help of bias monitoring. It gives them more profound insight into their training data and models to detect and mitigate bias and justify ML predictions.

Different ways to calculate bias

Disparity Impact ratio:

Disparity metrics can evaluate and contrast a model's performance across several groups as ratios or differnces.

• Disparity in model performance: These metrics sets determine the disparity (difference) in values of the selected performance metric across several subgroups.
• Disparity in selection rate: This measure includes the difference in selection rates between various subgroups.