Glossary · E-commerce ML
Algorithmic Fairness
also: ML fairness · fairness in machine learning · disparate impact · demographic parity · equalized odds
Definition
Algorithmic fairness is the study of disparate-impact, demographic-parity, equalized-odds, and related criteria for ML systems making consequential decisions. The Kleinberg/Chouldechova impossibility result (2016) proves that calibration, balance-for-positives, and balance-for-negatives cannot all hold simultaneously when base rates differ across groups.
Algorithmic fairness translates a moral question into a mathematical constraint set. Demographic parity requires equal positive-prediction rates across groups; equalized odds requires equal true-positive and false-positive rates; calibration requires that predicted probabilities mean the same thing across groups. The Kleinberg, Mullainathan, Raghavan 2016 and Chouldechova 2017 impossibility results show that these criteria are mutually incompatible when base rates differ, forcing a deliberate choice rather than a technical one. Production deployments often default to disparate-impact monitoring (Adverse Impact Ratio under the EEOC four-fifths rule) plus explicit bias-audit logging.
Essays on this concept
- E-commerce ML
Dynamic Pricing Under Demand Uncertainty: A Contextual Bandit Approach with Fairness Constraints
Airlines have done dynamic pricing for decades. E-commerce is catching up - but without the fairness constraints that prevent algorithms from charging different people different prices for the same product based on inferred willingness to pay.
- Pricing Strategy
Pricing Experimentation Without the Legal Risk: An Operator Framework for Defensible A/B Tests
Price A/B tests are not, by themselves, illegal. Most of the legal risk lies in how the cohorts are formed, what data is used, and what the team can show a regulator a year later. This is the framework that survives the question.
- E-commerce ML
Real-Time Fraud Detection at Checkout: A Streaming ML Pipeline Architecture with Sub-100ms Latency
You have 100 milliseconds to decide whether a transaction is fraudulent. In that window, you need to compute 200+ features from streaming data, run inference on a model trained on 1:1000 class imbalance, and return a score that balances revenue loss against customer friction.
- E-commerce ML
Personalized Promotion Optimization: Uplift Modeling to Identify Who Needs a Discount vs. Who Would Buy Anyway
70% of promotional spend goes to customers who would have purchased at full price. Uplift modeling identifies the 30% whose behavior actually changes with a discount, and ignores the rest. The math isn't complicated. The organizational willingness to stop blanket discounting is.
- E-commerce ML
Search Ranking as a Revenue Optimization Problem: Learning-to-Rank with Business Objective Regularization
E-commerce search is not Google search. When a user types 'running shoes,' the goal isn't to find the most relevant document, it's to surface the product most likely to be purchased at the highest margin. This reframes ranking as a constrained revenue optimization problem.
- Pricing Strategy
Dynamic Pricing Fairness Audits: A Practitioner Method for Pre-Launch and Continuous Review
Dynamic pricing systems can drift into discrimination without anyone in the team intending it. The audit method is borrowed from credit modeling, adapted for pricing CI/CD, and made boring enough to run every release.
Related concepts
Authoritative references