Add Fisher information fairness audit to Chapter 6 (Whose?)

[Rakshitha-Ireddi]

Authors

• Ireddi Rakshitha
• Yashwanth Devavarapu

Summary

This PR adds a Fisher information fairness audit to Chapter 6 (Whose?), giving a concrete way to check whether an elicitation policy allocates information fairly across groups.

Changes

• New subsection Fisher information fairness audit (#sec-fisher-fairness-audit) in the Elicitation section of Chapter 6, after the existing "Stratified Sampling vs. Active Learning" code example.
• Narrative: Explains that the same Fisher information used for active learning in Chapter 3 can be used to audit elicitation: we track cumulative Fisher information per group under different policies.
• Rasch model setup: Two groups (A/B), shared item bank, Gaussian prior per user. Fisher information for user (U) from item (j) is (\mathcal{I}_j(U) = p_j(U)(1 - p_j(U))) with (p_j(U) = \sigma(U + V_j)).
• Two policies:
  • Fisher-optimal (global): Each query is the (user, item) pair that maximizes Fisher information (information-maximizing).
  • Stratified Fisher: Allocate queries by group proportion (e.g. 60/40), then within each group choose the most informative item for that user.
• Pyodide code: Simulates both policies, plots cumulative Fisher information per group over time, and reports per-user information and disparity. Shows that Fisher-optimal can allocate more information to one group; stratified Fisher typically reduces disparity.
• Interpretation: Makes the efficiency–fairness tradeoff explicit and supports the design principle of stratifying when fairness is a goal.

Motivation

The book states that "Fisher information drives both active learning (Ch3) and fairness auditing (Ch6)." This addition makes that link operational: we use Fisher information as an audit metric (cumulative information per group) so practitioners can see whether their elicitation policy is fair in an information-theoretic sense.