We introduce a typed natural deduction system designed to formally verify the presence of bias in automatic labeling methods. The system relies on a ”data-as-terms” and ”labels-as-types” interpretation of formulae, with derivability contexts encoding probability distributions on training data. Bias is understood as the divergence that expected probabilistic labeling by a classifier trained on opaque data displays from the fairness constraints set by a transparent dataset.
Proof-checking Bias in Labeling Methods
F. A. D'Asaro
2022-01-01
Abstract
We introduce a typed natural deduction system designed to formally verify the presence of bias in automatic labeling methods. The system relies on a ”data-as-terms” and ”labels-as-types” interpretation of formulae, with derivability contexts encoding probability distributions on training data. Bias is understood as the divergence that expected probabilistic labeling by a classifier trained on opaque data displays from the fairness constraints set by a transparent dataset.
File in questo prodotto:
| File | Dimensione | Formato | |
|---|---|---|---|
|
paper1.pdf
accesso aperto
Licenza:
Creative commons
Dimensione
301.05 kB
Formato
Adobe PDF
|
301.05 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
