概述
It can persist even when data are collected and measured accurately, because the historical process itself was unequal. A model can therefore automate old patterns at greater scale while appearing objective.
深入探討
Historical bias is a source of harm that arises before or during model development because the data encode inequities that already exist. Suresh and Guttag’s machine-learning lifecycle framework distinguishes this from representation bias, measurement choices, aggregation, learning, evaluation, and deployment harms. A dataset can represent its source population accurately and still preserve unfair social structures. More data from the same process may make the pattern more statistically stable without making it fair. The mechanism depends on the target and data-generating process. If a hiring system predicts prior hiring decisions, it may treat past choices as a definition of merit. If a credit model uses variables shaped by segregation or exclusion, those variables can carry historical differences into future decisions even when race is omitted. In health care, Obermeyer and colleagues found that a widely used population-health algorithm used health-care cost as a proxy for need; because less money had historically been spent on Black patients with comparable illness, the score understated their need. The result was a target-design problem tied to structural differences, not simply an incorrectly measured feature. Historical bias differs from a data-coverage problem: adding more records of the same decisions may not address a discriminatory target. It also differs from label bias caused by annotator judgments, although both can be present together. Teams should ask what outcome the model is trained to predict, who had access to the opportunity in the past, and which institutional choices produced the labels. Mitigation starts by questioning the target and decision process. Use outcomes closer to the intended goal, examine whether historic labels are suitable, test subgroup impacts, and consider whether the institution should change the process rather than automate it. A fairness metric alone cannot decide what outcome is legitimate. Document the historical context, affected groups, and remaining uncertainty before deployment.
戰略影響
成本與預算
多年來，架構決策決定著效能和營運成本。
更明確的決策
技術教育幫助團隊選擇正確的堆疊，而不僅僅是最新的堆疊。
品質管控
更好的工程選擇可以減少生產中的可靠性事故。
The Future of Historical Bias in Machine Learning
Historical patterns change slowly and can re-enter models through refreshed labels, vendors, or workflow feedback. Reassess targets and outcomes after policy changes, not only after retraining. Keep an auditable account of the data-generating process and provide affected people a route to challenge decisions. Monitoring can reveal drift but cannot determine by itself whether a historical outcome is fair. Recheck subgroup effects when labels or policies change. Keep an appeal process so affected people can surface recurring errors. Review appeals for repeated patterns.
現實世界的實施
A résumé model trained on past hiring records learns to rank candidates from women’s colleges lower because the company hired few of them historically.
A loan model uses neighborhood history shaped by redlining and assigns worse terms to applicants from previously excluded areas.
A promotion model learns that part-time workers were rarely promoted and repeats that pattern without examining whether the historical decisions were fair.
A public-safety system trained on recorded incidents can mistake enforcement patterns for the true distribution of harmful behavior.
風險與防護欄
優化一項基準測試可以隱藏更廣泛的系統弱點。
基礎設施和維護成本常常被低估。
隨著系統變得更加複雜，安全性和可觀察性差距可能會擴大。
實施路線圖
在實施之前定義延遲、品質和成本目標。
在實際負載和資料條件下進行基準測試。
儀器監控錯誤、漂移和使用者影響。
在擴展之前準備回滾和事件回應路徑。
不斷探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Historical Bias in Machine Learning quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常見問題
What is Historical Bias in Machine Learning?
Historical bias occurs when patterns in past data reflect structural inequities or past discrimination, and a model learns to reproduce them. It can persist even when data are collected and measured accurately, because the historical process itself was unequal. A model can therefore automate old patterns at greater scale while appearing objective.
A model accurately learns patterns in prior hiring decisions that excluded a group. Which source of bias is most directly involved?
The model reproduces inequity already present in the historical decision process.
Why may collecting more records from the same historical process fail to remove historical bias?
More observations can stabilize patterns produced by an unequal process without correcting that process.
A health algorithm uses past spending to predict medical need. What risk did Obermeyer and colleagues identify?
The study found cost as a proxy for need led to lower scores for Black patients who were sicker at comparable risk scores.
Which question helps identify a potentially biased training target?
Historical bias analysis examines the outcomes and institutional choices that generated labels.
How does historical bias differ from representation bias?
A dataset can cover its source population yet still encode inequitable past outcomes.
繼續學習
相關指南
為此主題精選的更多指南