Introduction: What’s fairness?

The concept “fairness” was first proposed in about late 2000s, attracting a wide interest together with ethical AI in the last decade. In general, fairness can be divided into three categories: group fairness, individual fairness and conterfactual fairness, where the former two are statistical dependence measures and the latter is rooted in causal inference. So how do we define fairness?

Values of fairness in real-world applications

Fairness-Accuracy tradeoff

Regularization and inference with label constraints

  • Regularization with label constraints
  • On-training approach: Inference-based training
  • Learning plus inference

Achieve flexible fairness-accuracy tradeoff at test time

Learning with constraints

  • YOTO
  • YODO

Inference with constraints