Whose Polarization Is It? Auditing Fairness in Network Polarization Measures
We audit seven polarization measure families spanning the three network-operationalizable dimensions of polarization, introducing Network-Weighted Distance (NWD) to fill an unserved extremeness-separation combination.
ProofPaper ↗
Key points
- A polarization score is supposed to describe a community, not its surroundings.
- Yet when a minority community is embedded in a larger network, its measured polarization can rise or fall with the size and structure of the majority around it.
- We ask whether polarization measures are therefore fair to minorities and develop an audit built around two criteria: measurement invariance (MI), which asks whether otherwise equivalent majorities and minorities receive equivalent scores, and structural exogeneity (SE), which asks whether a minority's score changes when its surrounding majority changes.
- We trace this dependence to two separate channels: majority structure can alter the opinions that opinion-reading measures consume, and, on a different path, opinion-orthogonal structural change can affect the arithmetic of structural measures.
Sources (1)
- [1]Whose Polarization Is It? Auditing Fairness in Network Polarization MeasuresarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 09:47 PM
We audit seven polarization measure families spanning the three network-operationalizable dimensions of polarization, introducing Network-Weighted Distance (NWD) to fill an unserved extremeness-separation combination.
A polarization score is supposed to describe a community, not its surroundings.
Extractive summary: sentences quoted from the sources.