gatheluck / PaperReading

Notes about papers (in Japanese)
0 stars 0 forks source link

[2018] Mitigating Unwanted Biases with Adversarial Learning #49

Open gatheluck opened 4 years ago

gatheluck commented 4 years ago

論文リンク

https://arxiv.org/abs/1801.07593

公開日(yyyy/mm/dd)

2018/01/22 AAAI ACM Conference on AI, Ethics, and Society 2018

概要

データセットから分類器が学習してしまったバイアスを修正する手法(adversarial debiasig)を提案。予測精度の最大化と同時に、隠したいラベルの予測精度を最小化する。

gatheluck commented 4 years ago

【問題設定】

Y^ = f(X): predictor

通常はf(X)を通じてZの推定が可能になる。

gatheluck commented 4 years ago

debiasing という単語は以下の論文で提案されている。

gatheluck commented 4 years ago

【定義:Demographic Parity】 Y^はY^とZが独立のとき、Demographic Parityを満たすという。 つまり、P(Y^=y^)=P(Y^=y^|Z=z)となる。