ACM FAT*
f
f( )=A
f( )=A


f
f( )=A
f 

f( )=A


f
f
f( )=A
f
f 

X Y ̂YX
S = S =
X
Y
S
̂Y
ℙ{ ̂Y ∈ 𝒜|S = s} = ℙ{ ̂Y ∈ 𝒜|S = s′}
𝒜, s, s′
=
̂Y|S = ̂Y|S =
ℙ{ ̂Y ∈ 𝒜|Y = y, S = s} = ℙ{ ̂Y ∈ 𝒜|Y = y, S = s′}
𝒜, y, s, s′
Y ̂Y
Y = 1 ̂p
Y = 1 ̂p
ℙ{Y = 1| ̂p = p, S = s} = p
p, s
p
̂p = p|S =
x, x′
D( f(x), f(x′)) ≤ d(x, x′)
≈ ⟹
f : 𝒳 → Δ(𝒴)
f( )=A
f 

minf Err(f ) + ηUnfair(f )
minf Err(f ) Unfair(f ) ≤ η
Q
Q f
minQ 𝔼f∼Qℙ{f(X) ≠ Y} M𝔼f∼Q[μ(f )] ≤ c
𝔼{f(X)|S = 0} = 𝔼{f(X)}
𝔼{f(X)|S = 1} = 𝔼{f(X)}
minQ 𝔼f∼Qℙ{f(X) ≠ Y} M𝔼f∼Q[μ(f )] ≤ c
maxλ∈ℝK
+,∥λ∥≤B minQ 𝔼f∼Qℙ{f(X) ≠ Y} + λ⊤
(M𝔼f∼Q[μ(f )] − c)
minf ∑
n
i=1 (h(Xi)C1
i + (1 − h(Xi))C0
i )
λ Q
μ
Q
λ


g( )= z
f( )=A
zz
g( )
g( )
z
g( )
g( )
z
g( )
g( )
z
f( )=A
z
g( )
g( )
z
g( )=
f( )=A
d( )=z
z
z








minf Likelihood(f(X), Y) + ηI(f(X), S)
f( )=A
f( )=A
f( )=A


f( )=A


̂Y = 1
maxy,s (VC(ℱ) + ln(1/δ))/(nPy,s)
maxy,s ln(1/δ)/(nPy,s)
maxy,s (VC(ℱ) + ln(1/δ))/(nPy,s)
maxy,s ln(1/δ)/(nPy,s)
(y, s)


minθ 𝔼[ℓ0(X, θ)] 𝔼[ℓi(X, θ)] ≤ 0
ϵ
m
Rn(ℱ)
ϵ + Rn(ℱ) + ln(1/δ)/n
(m ln(1/ϵ) + ln(m/δ))/n
ϵ + Rn(ℱ) + ln(1/δ)/n
(m ln(1/ϵ) + ln(m/δ))/n
ϵ
m
Rn(ℱ)
̂Y = 1 h : 𝒳 → [0,1]
h ℓ0
ℙx,x′{|h(x) − h(x′)| > d(x, x′) + γ} ≤ α
(γ, α)
maxi,j max(0,|h(x) − h(x′)| − d(xi, xj)) ≤ γ
m = O(poly(1/ϵα,1/ϵγ,1/ϵ)) ϵ
(α + ϵα, γ + ϵγ) h




∑
T
t=1
r(t)
x(t)
1 , . . . , x(t)
K i
r(t)
= fi(x(t)
i )
x(t)
i(t)
r(t)
πi(t) > πj(t) fi(x(t)
i ) > fj(x(t)
j )
fi(x(t)
i )
K3
T ln(Tk/δ)
T4/5
K6/5
d3/5
∨ k3
ln(k/δ)
Ω( T) Ω( K3
ln(1/δ))


TKd ln(T)
πi(t) ≠ ℙ{i = arg maxj rj}
D(π(t)
i , π(t)
j ) ≤ ϵ1D(ri, rj) + ϵ2








(KT)2/3
1 − δ D(π(t)
i , π(t)
j ) ≤ 2D(ri, rj) + ϵ2
|πi(t) − πj(t)| ≤ d(x(t)
i , x(t)
j )
x(t)
i
π(t)
r(t)
O(t)



ϵ
r(t)
maxπ∈ΔK
∑
K
i=1
riπi |πi − πj | ≤ dij
K, d T
d T ln(T/δ)
K2
d2
ln(TKd)
K2
d2
ln(kdT/ϵ) + K3
ϵT + d T ln(T/δ)
K2
d2
ln(d/ϵ)
ϵ = 1/K3
T T
∑
∞
t=τ
γt−τ
r(t)
s(t)
a(t)
r(t)
ϵ
1/(1 − γ)
πi(t) > πj(t) fi(s(t)
i ) > fj(s(t)
j )


f
f
• [Hardt+16] Moritz Hardt, Eric Price, and Nathan Srebro.
Equality of Opportunity in Supervised Learning. In: NeurIPS,
pp. 3315-3323, 2016. https://arxiv.org/abs/1610.02413
• [Pleiss+17] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon
Kleinberg, and Kilian Q. Weinberger. On Fairness and
Calibration. In: NeurIPS, pp. 5680-5689, 2017. https://arxiv.org/
abs/1709.02012
• [Dwork+12] Cynthia Dwork, Moritz Hardt, Toniann
Pitassi, Omer Reingold, Rich Zemel. Fairness Through
Awareness. In: the 3rd innovations in theoretical computer
science conference, pp. 214-226, 2012. https://arxiv.org/abs/
1104.3913
• [Agarwal+18] Alekh Agarwal, Alina Beygelzimer, Miroslav
Dudík, John Langford, and Hanna Wallach. A Reductions
Approach to Fair Classification. In: ICML, PMLR 80, pp.
60-69, 2018. https://arxiv.org/abs/1803.02453
• [Agarwal+19] Alekh Agarwal, Miroslav Dudík, and Zhiwei
Steven Wu. Fair Regression: Quantitative Definitions and
Reduction-based Algorithms. In: ICML, PMLR 97, pp. 120-129,
2019. https://arxiv.org/abs/1905.12843
• [Zafar+13] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi,
and Cynthia Dwork. Learning Fair Representations. In: ICML,
PMLR 28, pp. 325-333, 2013.
• [Zhao+19] Han Zhao, Geoffrey J. Gordon. Inherent Tradeoffs in
Learning Fair Representations. In: NeurIPS, 2019, to appear.
https://arxiv.org/abs/1906.08386
• [Xie+16] Qizhe Xie, Zihang Dai, Yulun Du, Eduard
Hovy, Graham Neubig. Controllable Invariance through
Adversarial Feature Learning. In: NeurIPS, pp. 585-596, 2016.
https://arxiv.org/abs/1705.11122
• [Moyer+18] Daniel Moyer, Shuyang Gao, Rob
Brekelmans, Greg Ver Steeg, and Aram Galstyan. Invariant
Representations without Adversarial Training. In: NeurIPS, pp.
9084-9893, 2018. https://arxiv.org/abs/1805.09458
• [Woodworth+18] Blake Woodworth, Suriya Gunasekar, Mesrob
I. Ohannessian, Nathan Srebro. Learning Non-Discriminatory
Predictors. In: COLT, pp. 1920-1953, 2017. https://arxiv.org/abs/
1702.06081
• [Cotter+19] Andrew Cotter, Maya Gupta, Heinrich
Jiang, Nathan Srebro, Karthik Sridharan, Serena Wang, Blake
Woodworth, Seungil You. Training Well-Generalizing
Classifiers for Fairness Metrics and Other Data-Dependent
Constraints. In: ICML, PMLR 97, pp. 1397-1405, 2019. https://
arxiv.org/abs/1807.00028
• [Rothblum+18] Guy N. Rothblum, Gal Yona. Probably
Approximately Metric-Fair Learning. In: ICML, PMLR 80, pp.
5680-5688, 2018. https://arxiv.org/abs/1803.03242
• [Joseph+16] Matthew Joseph, Michael Kearns, Jamie
Morgenstern, Aaron Roth. Fairness in Learning: Classic and
Contextual Bandits. In: NeurIPS, pp. 325-333, 2016.
• [Liu+17] Yang Liu, Goran Radanovic, Christos
Dimitrakakis, Debmalya Mandal, David C. Parkes. Calibrated
Fairness in Bandits. In: 4th Workshop on Fairness,
Accountability, and Transparency in Machine Learning
(FATML), 2017. https://arxiv.org/abs/1707.01875
• [Gillen+18] Stephen Gillen, Christopher Jung, Michael
Kearns, Aaron Roth. Online Learning with an Unknown
Fairness Metric. In: NeurIPS, pp. 2600-2609, 2018. https://
arxiv.org/abs/1802.06936
• [Jabbari+17] Shahin Jabbari, Matthew Joseph, Michael Kearns, Jamie
Morgenstern, Aaron Roth. Fairness in Reinforcement Learning. In:
ICML, PMLR 70, pp. 1617-1626, 2017. https://arxiv.org/abs/1611.03071
• [Liu+18] Lydia T. Liu, Sarah Dean, Esther Rolf, Max
Simchowitz, Moritz Hardt. Delayed Impact of Fair Machine Learning.
In: ICML, PMLR 80, pp. 3150-3158, 2018. https://arxiv.org/abs/
1803.04383
• [Aivodji+19] Ulrich Aïvodji, Hiromi Arai, Olivier Fortineau, Sébastien
Gambs, Satoshi Hara, Alain Tapp. Fairwashing: the risk of
rationalization. In: ICML, 2019. https://arxiv.org/abs/1901.09749
• [Fukuchi+20] Kazuto Fukuchi, Satoshi Hara, Takanori Maehara. Faking
Fairness via Stealthily Biased Sampling. In: AAAI, Special Track on AI
for Social Impact (AISI), 2020, to appear. https://arxiv.org/abs/
1901.08291
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論
公平性を保証したAI/機械学習
アルゴリズムの最新理論

公平性を保証したAI/機械学習
アルゴリズムの最新理論

  • 1.
  • 3.
  • 8.
  • 9.
  • 10.
  • 11.
  • 12.
  • 13.
  • 14.
    X Y ̂YX S= S = X Y S ̂Y
  • 16.
    ℙ{ ̂Y ∈𝒜|S = s} = ℙ{ ̂Y ∈ 𝒜|S = s′} 𝒜, s, s′ = ̂Y|S = ̂Y|S =
  • 17.
    ℙ{ ̂Y ∈𝒜|Y = y, S = s} = ℙ{ ̂Y ∈ 𝒜|Y = y, S = s′} 𝒜, y, s, s′ Y ̂Y
  • 18.
    Y = 1̂p Y = 1 ̂p ℙ{Y = 1| ̂p = p, S = s} = p p, s p ̂p = p|S =
  • 19.
    x, x′ D( f(x),f(x′)) ≤ d(x, x′) ≈ ⟹ f : 𝒳 → Δ(𝒴)
  • 21.
  • 22.
    minf Err(f )+ ηUnfair(f ) minf Err(f ) Unfair(f ) ≤ η
  • 23.
    Q Q f minQ 𝔼f∼Qℙ{f(X)≠ Y} M𝔼f∼Q[μ(f )] ≤ c 𝔼{f(X)|S = 0} = 𝔼{f(X)} 𝔼{f(X)|S = 1} = 𝔼{f(X)}
  • 24.
    minQ 𝔼f∼Qℙ{f(X) ≠Y} M𝔼f∼Q[μ(f )] ≤ c maxλ∈ℝK +,∥λ∥≤B minQ 𝔼f∼Qℙ{f(X) ≠ Y} + λ⊤ (M𝔼f∼Q[μ(f )] − c)
  • 25.
    minf ∑ n i=1 (h(Xi)C1 i+ (1 − h(Xi))C0 i )
  • 26.
  • 27.
  • 28.
  • 29.
    g( )= z f()=A zz
  • 30.
  • 31.
  • 32.
  • 33.
  • 34.
    g( )= f( )=A d()=z z z 
 

  • 35.
  • 36.
  • 38.
  • 39.
  • 40.
  • 41.
  • 43.
  • 44.
    maxy,s (VC(ℱ) +ln(1/δ))/(nPy,s) maxy,s ln(1/δ)/(nPy,s)
  • 45.
    maxy,s (VC(ℱ) +ln(1/δ))/(nPy,s) maxy,s ln(1/δ)/(nPy,s) (y, s) 

  • 46.
    minθ 𝔼[ℓ0(X, θ)]𝔼[ℓi(X, θ)] ≤ 0
  • 47.
    ϵ m Rn(ℱ) ϵ + Rn(ℱ)+ ln(1/δ)/n (m ln(1/ϵ) + ln(m/δ))/n
  • 48.
    ϵ + Rn(ℱ)+ ln(1/δ)/n (m ln(1/ϵ) + ln(m/δ))/n ϵ m Rn(ℱ)
  • 50.
    ̂Y = 1h : 𝒳 → [0,1] h ℓ0 ℙx,x′{|h(x) − h(x′)| > d(x, x′) + γ} ≤ α (γ, α)
  • 51.
    maxi,j max(0,|h(x) −h(x′)| − d(xi, xj)) ≤ γ m = O(poly(1/ϵα,1/ϵγ,1/ϵ)) ϵ (α + ϵα, γ + ϵγ) h 

  • 52.
  • 53.
    ∑ T t=1 r(t) x(t) 1 , .. . , x(t) K i r(t) = fi(x(t) i ) x(t) i(t) r(t)
  • 54.
    πi(t) > πj(t)fi(x(t) i ) > fj(x(t) j ) fi(x(t) i )
  • 56.
    K3 T ln(Tk/δ) T4/5 K6/5 d3/5 ∨ k3 ln(k/δ) Ω(T) Ω( K3 ln(1/δ)) 
 TKd ln(T)
  • 57.
    πi(t) ≠ ℙ{i= arg maxj rj} D(π(t) i , π(t) j ) ≤ ϵ1D(ri, rj) + ϵ2
  • 58.
  • 59.
    
 
 (KT)2/3 1 − δD(π(t) i , π(t) j ) ≤ 2D(ri, rj) + ϵ2
  • 60.
    |πi(t) − πj(t)|≤ d(x(t) i , x(t) j ) x(t) i π(t) r(t) O(t) 
 
ϵ
  • 61.
  • 62.
    K, d T dT ln(T/δ) K2 d2 ln(TKd) K2 d2 ln(kdT/ϵ) + K3 ϵT + d T ln(T/δ) K2 d2 ln(d/ϵ) ϵ = 1/K3 T T
  • 63.
  • 64.
    ϵ 1/(1 − γ) πi(t)> πj(t) fi(s(t) i ) > fj(s(t) j )
  • 65.
  • 68.
  • 69.
  • 71.
    • [Hardt+16] MoritzHardt, Eric Price, and Nathan Srebro. Equality of Opportunity in Supervised Learning. In: NeurIPS, pp. 3315-3323, 2016. https://arxiv.org/abs/1610.02413 • [Pleiss+17] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q. Weinberger. On Fairness and Calibration. In: NeurIPS, pp. 5680-5689, 2017. https://arxiv.org/ abs/1709.02012 • [Dwork+12] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, Rich Zemel. Fairness Through Awareness. In: the 3rd innovations in theoretical computer science conference, pp. 214-226, 2012. https://arxiv.org/abs/ 1104.3913
  • 72.
    • [Agarwal+18] AlekhAgarwal, Alina Beygelzimer, Miroslav Dudík, John Langford, and Hanna Wallach. A Reductions Approach to Fair Classification. In: ICML, PMLR 80, pp. 60-69, 2018. https://arxiv.org/abs/1803.02453 • [Agarwal+19] Alekh Agarwal, Miroslav Dudík, and Zhiwei Steven Wu. Fair Regression: Quantitative Definitions and Reduction-based Algorithms. In: ICML, PMLR 97, pp. 120-129, 2019. https://arxiv.org/abs/1905.12843 • [Zafar+13] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning Fair Representations. In: ICML, PMLR 28, pp. 325-333, 2013.
  • 73.
    • [Zhao+19] HanZhao, Geoffrey J. Gordon. Inherent Tradeoffs in Learning Fair Representations. In: NeurIPS, 2019, to appear. https://arxiv.org/abs/1906.08386 • [Xie+16] Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, Graham Neubig. Controllable Invariance through Adversarial Feature Learning. In: NeurIPS, pp. 585-596, 2016. https://arxiv.org/abs/1705.11122 • [Moyer+18] Daniel Moyer, Shuyang Gao, Rob Brekelmans, Greg Ver Steeg, and Aram Galstyan. Invariant Representations without Adversarial Training. In: NeurIPS, pp. 9084-9893, 2018. https://arxiv.org/abs/1805.09458
  • 74.
    • [Woodworth+18] BlakeWoodworth, Suriya Gunasekar, Mesrob I. Ohannessian, Nathan Srebro. Learning Non-Discriminatory Predictors. In: COLT, pp. 1920-1953, 2017. https://arxiv.org/abs/ 1702.06081 • [Cotter+19] Andrew Cotter, Maya Gupta, Heinrich Jiang, Nathan Srebro, Karthik Sridharan, Serena Wang, Blake Woodworth, Seungil You. Training Well-Generalizing Classifiers for Fairness Metrics and Other Data-Dependent Constraints. In: ICML, PMLR 97, pp. 1397-1405, 2019. https:// arxiv.org/abs/1807.00028 • [Rothblum+18] Guy N. Rothblum, Gal Yona. Probably Approximately Metric-Fair Learning. In: ICML, PMLR 80, pp. 5680-5688, 2018. https://arxiv.org/abs/1803.03242
  • 75.
    • [Joseph+16] MatthewJoseph, Michael Kearns, Jamie Morgenstern, Aaron Roth. Fairness in Learning: Classic and Contextual Bandits. In: NeurIPS, pp. 325-333, 2016. • [Liu+17] Yang Liu, Goran Radanovic, Christos Dimitrakakis, Debmalya Mandal, David C. Parkes. Calibrated Fairness in Bandits. In: 4th Workshop on Fairness, Accountability, and Transparency in Machine Learning (FATML), 2017. https://arxiv.org/abs/1707.01875 • [Gillen+18] Stephen Gillen, Christopher Jung, Michael Kearns, Aaron Roth. Online Learning with an Unknown Fairness Metric. In: NeurIPS, pp. 2600-2609, 2018. https:// arxiv.org/abs/1802.06936
  • 76.
    • [Jabbari+17] ShahinJabbari, Matthew Joseph, Michael Kearns, Jamie Morgenstern, Aaron Roth. Fairness in Reinforcement Learning. In: ICML, PMLR 70, pp. 1617-1626, 2017. https://arxiv.org/abs/1611.03071 • [Liu+18] Lydia T. Liu, Sarah Dean, Esther Rolf, Max Simchowitz, Moritz Hardt. Delayed Impact of Fair Machine Learning. In: ICML, PMLR 80, pp. 3150-3158, 2018. https://arxiv.org/abs/ 1803.04383 • [Aivodji+19] Ulrich Aïvodji, Hiromi Arai, Olivier Fortineau, Sébastien Gambs, Satoshi Hara, Alain Tapp. Fairwashing: the risk of rationalization. In: ICML, 2019. https://arxiv.org/abs/1901.09749 • [Fukuchi+20] Kazuto Fukuchi, Satoshi Hara, Takanori Maehara. Faking Fairness via Stealthily Biased Sampling. In: AAAI, Special Track on AI for Social Impact (AISI), 2020, to appear. https://arxiv.org/abs/ 1901.08291