Production ML – Fairness

Model fairness is about assess how the output of the model can impact different subgroup of the population.

Tool:

Tensorflow Fairness Indicators

• For binary and multi-class classifiers
• Compares various performance metrics at slices of data (both features and label)
  • Accuracy
  • AUC ROC
  • AUC precision recall
  • Loss
  • Precision, Recall, etc (see Advanced Classification Metrics: Precision, Recall, and more)
• Built on top of Tensorflow Model Analysis

How to measure fairness?

• Positive/negative rate: examples labeled as positive/negative
  • Ex: in a loan approval classifier
    • Positive = approved loan
    • Negative = declined loan
    • Is the positive rate different between the male vs female population?
• Establish context for user groups
  • Require domain expert
  • Understand people’s social identity, social structure, and culture system
  • Talk to social scientists, linguists, anthropologists
  • Slice data widely and wisely
    • Race, ethnicity, gender, nationality, sexual orientation, income, disability status
• Evaluate metric across different thresholds
• Check low margin cases
• Compare TPR (true positive rate) vs FNR (false negative rate)
  • TPR – predicted true out of all actually true
  • FNR – predicted false out of all actually true
  • Compare TPR across different subgroups
  • Ex: need to ensure same positive rate across gender for qualified loan applicants