Fairness (machine learning)

• 1. Origins And Historical Context

  ›

  The year 1964 marked a turning point in American history when the U.S. Civil Rights Act passed into law. This legislation sparked intense debate within the scientific community about how to measure fairness and bias in decision-making processes. Researchers spent the next decade trying to define what it meant for an algorithm or human judge to be fair. By the end of the 1970s, these discussions had largely vanished from mainstream academic discourse. The competing notions of fairness left little room for clarity on which definition should take precedence. A sharp resurgence occurred only after 2016 when ProPublica released a report claiming that COMPAS software used in US courts was racially biased. This report reignited decades-old questions about whether automated systems could ever truly be impartial. Tech companies began releasing tools like IBM's Python libraries and Google's guidelines to detect and reduce bias. Facebook introduced Fairness Flow as a tool to identify discrimination in their AI programs. Critics argued these efforts remained insufficient because employees rarely used them and the tools did not cover all programs.

  1516

• 2. Language And Cultural Bias

  ›

  ChatGPT describes liberalism from an Anglo-American perspective while ignoring valid viewpoints from other cultures. When queried about political ideologies, the model emphasizes human rights and equality but omits aspects like opposing state intervention found in Vietnamese perspectives. It also fails to mention limitations of government power prevalent in Chinese thought. Japanese, Korean, French, and German corpora remain absent from its responses despite being multilingual chatbots. Luo et al. demonstrated that current large language models present Anglo-American views as truth systematically. They downplay non-English perspectives as irrelevant or noise during training processes. English-centric data causes this systematic deviation in sampling information. The true coverage of topics and views available in repositories gets distorted by this approach. A query about liberalism yields answers reflecting only one cultural framework. Other political perspectives embedded in different languages disappear entirely from generated text. This blindness extends beyond politics into gender roles where nurses are associated with women and engineers with men. LinkedIn profiles analyzed through Natural Language Processing methods reveal similar patterns of exclusion.

  1143

• 3. Real World Controversies

  ›

  In 2015 Google apologized after Photos mistakenly labeled a black couple as gorillas. Flickr's auto-tag feature later identified some black people as apes and animals instead of humans. An international beauty contest judged by AI in 2016 favored individuals with lighter skin due to biased training data. Three commercial gender classification algorithms tested in 2018 proved most accurate for light-skinned males but worst for dark-skinned females. Twitter's image cropping tool released in 2020 preferred lighter skinned faces over darker ones. DALL-E 2 creators explained in 2022 that their text-to-image model generated significantly stereotyped images based on traits like gender or race. Amazon used software to review job applications that penalized resumes containing the word women. Apple's credit card algorithm gave significantly higher limits to males than females even for couples sharing finances. The Markup reported mortgage-approval algorithms rejected non-white applicants more frequently in 2021. Eric Holder raised concerns in 2014 about risk assessment methods focusing on factors outside defendant control. ProPublica claimed black defendants were almost twice as likely to be incorrectly labeled higher risk compared to white defendants. Northpointe Inc disputed these findings claiming statistical errors before ProPublica refuted them again.

  1920212223

• 4. Mathematical Fairness Criteria

  ›

  Independence requires sensitive characteristics to be statistically independent of predictions made by classifiers. This means classification rates must equal across different groups regarding sensitive attributes. Mutual information between random variables provides another way to express this concept mathematically. Separation demands sensitivity independence given the target value itself. True positive and false positive rates become equal for every value of sensitive characteristics under this definition. Sufficiency states probability of being in each group equals for two individuals with different sensitive characteristics if predicted same group. Total fairness occurs when all three criteria hold simultaneously but remains impossible except in rhetorical cases. Confusion matrices display relationships between true positives, false negatives, and other metrics used to measure accuracy. Positive predicted value represents fraction of correct positive predictions out of all positive ones. False discovery rate measures erroneous positive predictions among total positive outputs. Negative predicted value indicates correct negative predictions relative to all negative cases. False omission rate captures probability of erroneous negative predictions within negative outcomes. True positive rate reflects correctly classified positive subjects while false negative rate shows incorrect classifications. True negative rate and false positive rate complete the set of standard evaluation metrics. Equalized odds combines equality of true positive and false positive rates across protected groups. Conditional use accuracy equality ensures equal precision and negative predictive values for both sides.

  414254

• 5. Mitigation Strategies

  ›

  Preprocessing algorithms remove information about dataset variables that might result in unfair decisions before training begins. Mapping individuals into intermediate representations hides protected attributes while maintaining maximum possible accuracy. Reweighing assigns weights to dataset points so weighted discrimination becomes zero regarding designated groups. Inprocessing adds constraints directly to optimization objectives during software training phases. Adversarial debiasing trains two classifiers simultaneously through gradient-based methods like gradient descent. The predictor tries minimizing loss function while maximizing adversary's failure to predict sensitive variable. Postprocessing adjusts thresholds applied to classifier scores after initial prediction generation occurs. Plotting true positive against false negative rates at various threshold settings helps find fair operating points. Reject option based classification labels unclear instances differently depending on whether they belong to deprived or privileged groups. These approaches attempt correcting bias at data level, model training stage, or final output phase respectively. Preprocessed datasets allow reuse across any machine learning task without modifying underlying classifiers. Inprocessing often yields better results in accuracy and fairness compared to preprocessing techniques alone. Adversarial learning improves demographic parity significantly when trained with an active adversary component present.

  515253