Debiaser for Multiple Variables (DEMV)

DEMV is a Debiaser for Multiple Variables that aims to increase Fairness in any given dataset, both binary and categorical, with one or more sensitive variables, while keeping the accuracy of the classifier as high as possible. The main idea behind the proposed method is that to enhance the classifier’s fairness during pre-processing effectively is necessary to consider all possible combinations of the values of the sensitive variables and the label’s values for the definition of the so-called sensitive groups. We approach the problem by recursively identifying all the possible groups given by combining all the values of the sensible variables with the belonging label (class). Next, for each group, we compute its expected (π‘Šπ‘’π‘₯𝑝) and observed (π‘Šπ‘œπ‘π‘ ) sizes and look at the ratio among these two values. If π‘Šπ‘’π‘₯𝑝/π‘Šπ‘œπ‘π‘  = 1, it implies that the group is fully balanced. Otherwise, if the ratio is less than one, the group size is larger than expected, so we must remove an element from the considered group accordingly to a chosen deletion strategy. Finally, if the ratio is greater than one, the group is smaller than expected, so we have to add another item accordingly to a generation strategy. For each group, we recursively repeat this balancing operation until π‘Šπ‘’π‘₯𝑝/π‘Šπ‘œπ‘π‘  converge to one. It is worth noting that, in order to keep a high level of accuracy, the new items added to a group should be coherent in their values with the already existing ones.

Data and Resources
To access the resources you must log in
  • Tutorial Notebookipynb

    A jupyter notebook showing the main functionalities of DEMV

    The resource: 'Tutorial Notebook' is not accessible as guest user. You must login to access it!
Additional Info
Field Value
Accessibility Both
AccessibilityMode OnLine Access
Availability On-Line
Basic rights Temporary download of a single copy only
CreationDate 2024-02-21 14:40
Creator d'Aloisio, Giordano,
Field/Scope of use Any use
Group Social Impact of AI and explainable ML
Owner d'Aloisio, Giordano,
SoBigData Node SoBigData EU
SoBigData Node SoBigData IT
Sublicense rights No
Territory of use World Wide
Thematic Cluster Other
system:type Method
Management Info
Field Value
Author d'Aloisio Giordano
Maintainer d'Aloisio Giordano
Version 1
Last Updated 21 February 2024, 16:05 (CET)
Created 21 February 2024, 15:43 (CET)