Skip to content

Fairness Assessment

Fairness Assessment

Overview

This report summarizes the existing research on fairness assessment of machine learning models, covering: key types of fairness definitions, fairness assessment metrics and challenges of fairness assessment.

Fairness in Machine Learning

1. Group Fairness

Group fairness aims to ensure equitable treatment across different demographic or protected groups (e.g., gender, race, age). Rather than focusing on individuals, group fairness metrics assess whether statistical measures of model performance are balanced across groups.

2. Individual Fairness

Individual fairness is based on the principle that similar individuals should be treated similarly. Tis principle deals with the comparison of single individuals rather than focusing on groups of people sharing some characteristics.

3. Observational vs. Causal Fairness

  • Observational:

Based on statistical relationships in the data. Does not attempt to model or infer the underlying causal mechanisms that generated the data. E.g. Demographic Parity (Equal acceptance rates across groups), Equality of Odds (Equal false positive and false negative rates across groups), Predictive Parity (Equal precision across groups).

It's easy to compute from data and does not require assumptions about how features or outcomes are generated, however, it can be misleading if the data itself is biased (e.g., due to historical discrimination) and it may enforce fairness in a way that ignores the reasons behind observed disparities. For example, enforcing demographic parity might require treating different groups unequally to equalize outcomes, even if those differences are due to legitimate factors.

  • Causal:

Based on counterfactual reasoning and structural causal models (SCMs).

​ Causal fairness metrics rely on understanding the causal relationships between variables. They aim to answer questions like: “Would this individual have received the same outcome if they had a different gender or race?”. E.g. Counterfactual Fairness (CFF): An individual should receive the same prediction in a counterfactual world where only their sensitive attribute (e.g., gender) is changed.

​ It requires a causal model, and distinguish between fair and unfair sources of disparity but it's hard to implement since it requires domain knowledge and assumptions about causal structure. Additionally counterfactuals are not directly observable, so validation is challenging.

Fairness Assessment Metrics

Metric Category Description
Demographic/Statistical Parity* Group Fairness Each sensitive group has the same probability of being included in the positive predicted class.
Equal Opportunity* Group Fairness True positive rates should be equal across groups.
Equalized Odds* Group Fairness True positive rate and false positive rate should be equal across groups.
Predictive Parity* Group Fairness Same precision across different groups.
Calibration Group Fairness For a given predicted score, the actual outcome should be the same across groups.
Counterfactual Fairness (Flip-rate is an empirical proxy for counterfactual fairness) Individual Fairness A prediction for an individual should remain the same if we were to change a sensitive attribute while keeping everything else constant.
Fairness Through Awareness Individual Fairness If two individuals are similar in relevant ways (according to a task-specific similarity metric), then a fair model should give them similar predictions.
Consistency Individual Fairness For similar individuals (for features excluding sensitive attributes) predictions should also be similar (often calculated via KNN)

*The fairness assessment metrics we're going to look into for the MCO model.

Challenges of Fairness Assessment in Machine Learning

  • Many group fairness metrics are mutually incompatible—you cannot satisfy all of them simultaneously unless under trivial conditions. E.g.
  • You cannot simultaneously satisfy Demographic Parity and Equality of Odds unless base rates are equal.
  • Predictive Parity and Equality of Odds are incompatible if the target variable is correlated with the sensitive attribute.
  • For individual fairness assessment like FTA (Fairness Though Awareness), task-specific similarity metric is subjective and requires domain expertise. In some domains, defining similarity is ethically and socially complex. Also it's computationally complex as we need to compare all pairs of individuals in the dataset which leads to quadratic time complexity.

Reference