eddings / fairsight

A visual analytic system for fair data-driven decision making

FairSight is a viable fair decision making system to assist decision makers in achieving fair decision making through the machine learning workflow.

Reference

Ahn, Y., & Lin, Y. R. (2019). Fairsight: Visual analytics for fairness in decision making. IEEE transactions on visualization and computer graphics (TVCG), 26(1), 1086-1095.

Specification

• React: Frontend framework for rendering and communicating with data
• django: Python-based backend framework for serving API of data and running machine learning work
• scss: The stylesheet grammar for more flexible structure
• d3.js: Javascript-based visualization library

FairDM

FairSight is developed on top of FairDM, a general fair decision making framework. Our framework is a model-agnostic framework with its goal to provide a fairness pipeline to guide the examination of fairness at each step (from input to output) in the workflow.

System

(a) Generator: The workflow of FairSight starts with setting up inputs including the sensitive attribute and protected group.
(b) Ranking View: After running a model, the ranking outcome and measures are represented.
(c) Global Inspection View: Visualizes the two spaces and the mapping process of Individual and Group fairness provided in the separate tap.
(d) Local Inspection View: When an individual is hovered, Local Inspection View provides the instance- and group-level exploration.
(e) Feature Inspection View: Users can investigate the feature distortion and feature perturbation to identify features as the possible source of bias.
(f) Ranking List View: All generated ranking outcomes are listed and plotted.