Slide 1
Slide 1 text
Fairness in Learning:
Classic and Contextual
Bandits
Jamie Morgenstern
joint w. Mathew Joseph, Michael Kearns, Aaron Roth
University of Pennsylvania
1
Slide 2
Slide 2 text
Automated decisions of
consequence
2
Hiring Lending Policing/
sentencing/
parole
[Miller, 2015],[Byrnes, 2016]
[Rudin, 2013], [Barry-Jester et al., 2015]
Slide 3
Slide 3 text
No content
Slide 4
Slide 4 text
Each individual has inherent ‘quality’
Slide 5
Slide 5 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
Slide 6
Slide 6 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
Slide 7
Slide 7 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
Slide 8
Slide 8 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Slide 9
Slide 9 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Slide 10
Slide 10 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Observe features, not qualities directly,
Slide 11
Slide 11 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Observe features, not qualities directly,
must learn relationship
Slide 12
Slide 12 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Observe features, not qualities directly,
must learn relationship
(observe loan application)
Slide 13
Slide 13 text
Each individual has inherent ‘quality’
(expected revenue for giving a loan)
entitling them to access to a resource
(high-revenue individuals deserve loans)
Observe features, not qualities directly,
must learn relationship
(observe loan application)
(assume different for different groups!)
Slide 14
Slide 14 text
A source of bias in ML
4
Slide 15
Slide 15 text
A source of bias in ML
4
• Data feedback loops: only observe/update estimates
for individuals current model believes are high-
qualilty
Slide 16
Slide 16 text
A source of bias in ML
4
• Data feedback loops: only observe/update estimates
for individuals current model believes are high-
qualilty
Slide 17
Slide 17 text
We study
5
Slide 18
Slide 18 text
We study
5
A new notion of fairness:
high-quality individuals must be
treated as well as lower-quality
individuals
Slide 19
Slide 19 text
We study
5
A new notion of fairness:
high-quality individuals must be
treated as well as lower-quality
individuals
And the “cost” of this constraint
wrt learning rate R(T)
(regret minimization)
Slide 20
Slide 20 text
We study
5
A new notion of fairness:
high-quality individuals must be
treated as well as lower-quality
individuals
And the “cost” of this constraint
wrt learning rate R(T)
(regret minimization)
Fair learning
rather than
finding a
fair model
Slide 21
Slide 21 text
Assumptions
k groups
each group has
a function mapping features to ’qualities’
(initially unknown, belonging to C)
(can be different for different groups)
Slide 22
Slide 22 text
Information/decision model
Each day t [T]
Observe feature vector from each group
Choose one individual based on features
Observe noisy estimate of quality of chosen
Goal: maximize expected average quality
Slide 23
Slide 23 text
Fairness Definition
8
Slide 24
Slide 24 text
Fairness Definition
8
An algorithm A( ) is fair if, for all (0, 1]
Slide 25
Slide 25 text
Fairness Definition
8
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
Slide 26
Slide 26 text
Fairness Definition
8
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
For any sequence x1, . . . , xT
Slide 27
Slide 27 text
Fairness Definition
8
for all rounds
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
For any sequence x1, . . . , xT
Slide 28
Slide 28 text
Fairness Definition
8
for all rounds
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
and all pairs of groups i, j
For any sequence x1, . . . , xT
Slide 29
Slide 29 text
Fairness Definition
8
for all rounds
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
and all pairs of groups i, j
For any sequence x1, . . . , xT
If E[quality of i at t] ≥ E[quality of j at t] then
Slide 30
Slide 30 text
Fairness Definition
8
for all rounds
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
and all pairs of groups i, j
For any sequence x1, . . . , xT
If E[quality of i at t] ≥ E[quality of j at t] then
A favors i over j in round t
Slide 31
Slide 31 text
Fairness Definition
8
for all rounds
with probability 1
An algorithm A( ) is fair if, for all (0, 1]
and all pairs of groups i, j
For any sequence x1, . . . , xT
If E[quality of i at t] ≥ E[quality of j at t] then
P A chooses i at t |
x1, . . . , xt
P A chooses j at t |
x1, . . . , xt
Slide 32
Slide 32 text
Separation between
fair and unfair learning
without features
9
Slide 33
Slide 33 text
Separation between
fair and unfair learning
without features
9
Theorem 1
Slide 34
Slide 34 text
Separation between
fair and unfair learning
without features
9
For any fair algo, R(T) = ˜ k3T .
Theorem 1
Slide 35
Slide 35 text
Separation between
fair and unfair learning
without features
9
Without fairness one can achieve regret
For any fair algo, R(T) = ˜ k3T .
Theorem 1
Slide 36
Slide 36 text
Separation between
fair and unfair learning
without features
9
Without fairness one can achieve regret
˜
O( kT)
For any fair algo, R(T) = ˜ k3T .
Theorem 1
Slide 37
Slide 37 text
Feature-based fair learning
10
[Strehl and Littman, 2008]
[ Li et al., 2011]
Slide 38
Slide 38 text
Feature-based fair learning
10
[Strehl and Littman, 2008]
[ Li et al., 2011]
Theorem 2 implications:
Slide 39
Slide 39 text
Feature-based fair learning
10
[Strehl and Littman, 2008]
[ Li et al., 2011]
Theorem 2 implications:
There is a fair algorithm for d-dimensional linear
mappings from features to qualities w. regret
R(T) = O T1 c · poly(k, d, ln 1 )
Slide 40
Slide 40 text
Feature-based fair learning
10
[Strehl and Littman, 2008]
[ Li et al., 2011]
Theorem 2 implications:
There is a fair algorithm for d-dimensional linear
mappings from features to qualities w. regret
R(T) = O T1 c · poly(k, d, ln 1 )
Any fair algorithm for d-dimensional conjunction
mappings must have regret R(T) = Ω 2d
Slide 41
Slide 41 text
Fairness through Datamining
- Pedreshi et al, ’08
- HDF ’13, HDFMB ’11, ZKC ’11, ….
“Group” Fairness
- CV ’10, FKL ’16, JL’15, KC ’11,
KKZ ’12, KAAS ’12…
“Individual” Fairness
- Dwork et al, 2012
- Johnson, Foster, Stine ’16
- Hardt, Price, Srebro ’16
- Kleinberg, Mullainathan, Raghavan ’16
11
Related Work
Slide 42
Slide 42 text
Fairness through Datamining
- Pedreshi et al, ’08
- HDF ’13, HDFMB ’11, ZKC ’11, ….
“Group” Fairness
- CV ’10, FKL ’16, JL’15, KC ’11,
KKZ ’12, KAAS ’12…
“Individual” Fairness
- Dwork et al, 2012
- Johnson, Foster, Stine ’16
- Hardt, Price, Srebro ’16
- Kleinberg, Mullainathan, Raghavan ’16
11
Related Work
Our work
focuses on fair
learning rather
than finding a
fair model
Slide 43
Slide 43 text
Conclusions
12
Slide 44
Slide 44 text
Conclusions
12
New notion of fairness:
higher-quality ⇒ better treatment
Slide 45
Slide 45 text
Conclusions
12
Must be *confident* about relative qualities
before preferential treatment ensues
New notion of fairness:
higher-quality ⇒ better treatment
Slide 46
Slide 46 text
Conclusions
12
There’s a cost to fairness
Must be *confident* about relative qualities
before preferential treatment ensues
New notion of fairness:
higher-quality ⇒ better treatment
Slide 47
Slide 47 text
Conclusions
12
There’s a cost to fairness
in some cases, this cost is mild, in others, great
Must be *confident* about relative qualities
before preferential treatment ensues
New notion of fairness:
higher-quality ⇒ better treatment
Slide 48
Slide 48 text
Conclusions
12
There’s a cost to fairness
in some cases, this cost is mild, in others, great
Must be *confident* about relative qualities
before preferential treatment ensues
New notion of fairness:
higher-quality ⇒ better treatment
Thanks!