###### tags: `On_Boarding Week` # OnBoradin week (day 2) ## Technical Skills: ### Salvatore Ruggieri: Discrimination and fairness * KDDLAB - University of Pisa, Iatly * Bias in AI and Risks * Cognitive BIAS Codex (figure) * Science Journal: paper * Datasets of CyCAT or OSAI by AI4EU * Impact of Bias on several aspect of our lives: Loss of Opportunity-Economic Loss, harms both to individual and society - social detriment - loss of Liberty ### Causes of Risk: * Bias: deviation from the truth * Internal Validity * Bias of Input * Bias of the Algorithm * Bias in the process * Examples: Bias of Input: Confounding ----- it is Simpons Paradox * Examples: Bias of Alg: Spurious correlations ----- * How Big Data is Unfair * Ref: Sources and types of Bias * Another example: Bias difficult to discover like the example of number of red cards given to black soccer players over 29 referrees * Definition of Discrimination: an example is the ZIP code correlated to the race (redline) ## Economic models of Discrimination * Statistical discrimination * Taste-based Discrimination * Methods: observational data * Regression model * Salary discrimination in major League Baseball. Labor Economics 2011 * example: Getting a higher COMPAS score * Methods: auditing/situation testing * example: Evidence from Ebay ## Discrimination: CS overview * discrimination discovery from data - data mining techniques * Discrimination measures * Discrimination measures: degree of discrimination suffered * Risk difference - ... * Jury selection example from men and women * software tools: Aequitas * alph-protection ... * implementation of alpha-protection FairTest * Indirect discrimination discovery * Lipschitz condition * Software tools: AdFisher ## Second part of speech : Fariness through decision making * data sanitization * Instance reweighting * * Promotion/demotion of decisions * Feature distribution distortion * .. * Fair Classification Learning * Fariness through awareness * Accuracy of classifiers: * extend clasic measures * Accuracy is an estimate of probability of correct predictions. * Example of COMPAS: prediction errors * FPR - TPR - NR * Again COMPAS: prediction errors * Fairness Measures: * Accuracy equality * predictive equality * Equal opportunity * . * . * conditional statistical parity * references * List of Fair algorithms * Github comparison between algorithms * Fairness and Causlaity: #### Post processing models * IBM post-porcessing #### tools & References * Text book ongoing * Facct conf * other conferences