Page 57 - Read Online
P. 57

Su et al. Intell Robot 2022;2(3):244­74  I http://dx.doi.org/10.20517/ir.2022.17    Page 250

                                            Table 1. Typical causality-based fairness notions
                 Type    Fairness notion         Formulation               Description
                                                                           The causal effects of the value change of
                                                                           the sensitive attribute    from    to    on
                                                                                                 −
                                                                                              +
                         Total Effect  [10]      |  (  |   =    ) −   (  |   =    ) | ≤     decision    =   , where the intervention is
                                                       +
                                                               −
                                                                           transmitted along all causal paths and is within
                                                                           the fair threshold   
                                                                           The difference between the distribution of
                          Effect of treatment on  |  (      + |   =    ) −   (      − |   =    ) | ≤        =    had    been    and that of    =   
                                                                                      +
                                                        −
                                                                 −
                          the treated  [10]                                had    been    , given that    had been
                                                                                   −
                                                                           observed to be    −
                 Group                                                     The causal effects of the value change of
                                                                           the sensitive attribute    from    to    on
                                                                                                 −
                                                                                              +
                         Path-specific fairness  [26,27]  |  (  |    (   |  ,    | ¯  )) −   (  |    (   )) | ≤   
                                                       +
                                                                    −
                                                          −
                                                                           decision    =    along specific causal
                                                                           paths, is within the fair threshold   
                                                                           It is satisfied when there is no directed path
                         No unresolved discrimination  [28]  -             from sensitive attribute    to outcome   
                                                                           allowed, except through a resolving variable
                                                                           If it is satisfied, there is no path from the
                         No proxy discrimination  [28]    (   |    (   =    0 )) =   (   |    (   =    1 ))  sensitive attribute    to the outcome    blocked
                                                  ∀   0 ,    1 ∈       (  )
                                                                           by a proxy variable
                                                                           An outcome    achieves counterfactual fairness
                                                                           towards an individual    (i.e., O = o) if the
                                                            −
                         Counterfactual fairness  [11]  |  (      + |O = o,    =    )−  probability of    =    for such individual    is
                                                    (      − |O = o,    =    ) | ≤     the same as the probability of    =    for the
                                                            −
                                                                           same individual, who belongs to a different
                                                                           sensitive group
                 Individual                                                It is based on situation testing where the
                         Individual direct discrimination  [29]    (  ,    ) =  ∑ |X|  |     (      ,    ) ·      (      ,    ) |  causal reasoning is used to define the
                                                                      ′
                                                    ′
                                                               ′
                                                         =1             
                                                                           distance function   (  ,    ) ′
                                                                           It detects discrimination by comparing the
                                                                           effort required to reach the same level of
                         Equality of effort  [30]  Ψ    + (  ) = Ψ    − (  )  outcome of individuals from advantaged
                                                                      
                                                  where Ψ    + (  ) =                ∈   E[      + ] ⩾     and disadvantaged groups who are similar to
                                                                           the target individual
                                                                           It is a general fairness formalization for
                                                                           representing various causality-based
                 Hybrid  PC-fairness  [31]       |  ( ˆ      + |  ,   − | ¯   |o) −   ( ˆ      − |o) | ≤   
                                                                           fairness notions, which is achieved
                                                                           by differently tuning its parameters
                                         Statistical-based notions
                                                     Statistical parity
                                                     Equalized odds
                                        Association
                                                     Predictive parity
                                                         ......
                                                                                   Total effects
                                                     Total effects
                                                                        Group-based  Path-specific fairness
                      Fairness notions  Intervention  Path-specific fairness       ······
                                                      ......
                                                                                   Based on   Equality of effort
                                                     Counterfactual fairness       similarity
                                                                         Individual  individuals ······
                                       Counterfactual  PC-Fairness       -based           Counterfactual
                                                      ......                       Based on   fairness
                                                                                   predicted
                                          Causality-based notions                  outcome  ······
                                             Figure 4. The categorization of fairness notions.
               In the real world, the focus of different machine learning tasks is different, and thus, various causality-based
               fairness notions are proposed to detect discrimination in different scenarios. This section introduces several
               representative causality-based fairness measurements that quantify fairness from the perspective of groups or
               individuals, respectively. Without loss of generality, assume that the sensitive attribute    and the outcome
               of the automated decision making    are binary variables where    =    denotes the advantaged group (e.g.,
                                                                           +
   52   53   54   55   56   57   58   59   60   61   62