This tutorial introduces causal modeling, presents a causal-modeling framework for anti-discrimination learning, and surveys recent research in the area.

Anti-discrimination learning addresses discrimination in data mining and machine learning. Discrimination discovery analyzes historical decision records to identify discriminatory practices, while discrimination prevention modifies biased data or predictive algorithms. Because discrimination is causal, demonstrating it requires evidence of a causal relationship rather than an association. Although association does not imply causation, this distinction has received limited attention in the literature. The tutorial surveys association-based approaches and their limitations, introduces causal modeling and a causal-modeling framework for anti-discrimination learning, and concludes with directions for future research.

Outline

• Introduction
  • Context
  • Literature and resources
• Correlation-based Anti-Discrimination Learning
  • Measuring discrimination:
    • fairness through unawareness,
    • disparate impact, individual fairness,
    • statistical parity,
    • equality of opportunity,
    • calibration,
    • conditional discrimination,
    • alpha discrimination,
    • multi-factor interaction,
    • belief,
    • preference.
  • Removing discrimination:
    • pre-processing (data modification, fair data representation, fair data generation),
    • in-processing (regularization, explicit constraints),
    • post-processing.
  • From correlation to causation: discrimination as causal effect, literature, motivating examples
• Causal modeling background
  • From statistics to causal modeling
  • Structural causal model and causal graph: Markovian model, conditional independence, d-separation, factorization formula
  • Causal inference: intervention and do-operator, truncated factorization formula, path-specific effect, identifiability and “kite” structure, counterfactual analysis
• Causal modeling-based anti-discrimination learning
  • Direct and indirect discrimination
  • Counterfactual fairness
  • Data discrimination vs. model discrimination
  • Other works on causal modeling-based anti-discrimination learning
• Challenges and directions for future research
  • Summary of existing works
  • Challenges: non-identifiability of path-specific effects, causal modeling for mixed-type variables, Semi-Markovian and ADMG, uncertain causal models, group/individual-level indirect discrimination
  • Future directions: building discrimination-free predictors, discrimination in tasks beyond classification: ranking and recommendation, Generative Adversarial Network (GAN), dynamic data and time series, text and image