New paper: 'On the Apparent Conflict Between Individual and Group Fairness' accepted at @fatconference, (now up on arxiv.org/abs/1912.06883). The paper addresses a distinction drawn between two broad approaches to measuring fairness in machine learning
'individual fairness' measures compare individuals, e.g. people 'similar' according to some task-relevant metric should get the same outcome. 'Group fairness' measures compare protected groups (e.g. gender, race, age) for differences in errors/outcomes/calibration/etc.
As typically presented, they reflect different + conflicting normative principles: group fairness seems to ensure egalitarian equality, whereas individual fairness seems to ensure consistency (the more/less qualified an individual, the better/worse their outcome).
But (I argue) the normative conflict only aligns that way if you pick a version of individual fairness which assumes there *isn't* any structural discrimination, and a version of group fairness which assumes there *is*.
But some versions of group fairness preserve consistency and ignore structural discrimination (e.g. equal calibration), and conversely, some versions of individual fairness can factor structural discrimination into the metric, adjusting scores so otherwise 'different' people...
... are 'similar' once disadvantage taken into account (e.g. in Dwork etal '12 re: SATs in college admissions). So conflict isnt between individual vs group per se; its about different worldviews, which can be reflected in variants from either kind of measure
Finally: individual fairness may appear to protect what 🇩🇪 constitution calls Einzelfallgerechtigkeit (individual justice). But even individually-fair ML is not individually just in this sense, because it still generalises between individuals who share same point in feature space
• • •
Missing some Tweet in this thread? You can try to
force a refresh
