このページは http://www.slideshare.net/shima__shima/fairnessaware-classifier-with-prejudice-remover-regularizer の内容を掲載しています。

掲載を希望されないスライド著者の方は、こちらよりご連絡下さい。

約4年前 (2012/09/25)にアップロードinテクノロジー

Fairness-aware Classifier with Prejudice Remover Regularizer

Proceedings of the European Conferen...

Fairness-aware Classifier with Prejudice Remover Regularizer

Proceedings of the European Conference on Machine Learning and Principles of Knowledge Discovery in Databases (ECMLPKDD), Part II, pp.35-50 (2012)

Article @ Official Site: http://dx.doi.org/10.1007/978-3-642-33486-3_3

Article @ Personal Site: http://www.kamishima.net/archive/2012-p-ecmlpkdd-print.pdf

Handnote: http://www.kamishima.net/archive/2012-p-ecmlpkdd-HN.pdf

Program codes : http://www.kamishima.net/fadm/

Conference Homepage: http://www.ecmlpkdd2012.net/

Abstract:

With the spread of data mining technologies and the accumulation of social data, such technologies and data are being used for determinations that seriously affect individuals’ lives. For example, credit scoring is frequently determined based on the records of past credit data together with statistical prediction techniques. Needless to say, such determinations must be nondiscriminatory and fair in sensitive features, such as race, gender, religion, and so on. Several researchers have recently begun to attempt the development of analysis techniques that are aware of social fairness or discrimination. They have shown that simply avoiding the use of sensitive features is insufficient for eliminating biases in determinations, due to the indirect influence of sensitive information. In this paper, we first discuss three causes of unfairness in machine learning. We then propose a regularization approach that is applicable to any prediction algorithm with probabilistic discriminative models. We further apply this approach to logistic regression and empirically show its effectiveness and efficiency.

- Fairness-aware Learning through Regularization Approach5年弱前 by Toshihiro Kamishima
- Future Directions of Fairness-Aware Data Mining: Recommendation, Causality, and Theoretical Aspects1年以上前 by Toshihiro Kamishima
- Correcting Popularity Bias by Enhancing Recommendation Neutrality約2年前 by Toshihiro Kamishima

- Fairness-aware Classifier

with Prejudice Remover Regularizer

Toshihiro Kamishima*, Shotaro Akaho*, Hideki Asoh*, and Jun Sakuma†

*National Institute of Advanced Industrial Science and Technology (AIST), Japan

†University of Tsukuba, Japan; and Japan Science and Technology Agency

ECMLPKDD 2012 (22nd ECML and 15th PKDD Conference)

@ Bristol, United Kingdom, Sep. 24-28, 2012

START

1 - Overview

Fairness-aware Data Mining

data analysis taking into account potential issues of

fairness, discrimination, neutrality, or independence

fairness-aware classification, regression, or clustering

detection of unfair events from databases

fairness-aware data publication

Examples of Fairness-aware Data Mining Applications

exclude the influence of sensitive information, such as gender or

race, from decisions or predictions

recommendations enhancing their neutrality

data analysis while excluding the influence of uninteresting

information

2 - Outline

Applications

fairness-aware data mining applications

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, fairness-aware data publication

Conclusion

3 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

4 - Elimination of Discrimination

Due to the spread of data mining technologies...

Data mining is being increasingly applied for serious decisions

ex. credit scoring, insurance rating, employment application

Accumulation of massive data enables to reveal personal information

ex. demographics can be inferred from the behavior on the Web

excluding the influence of these socially sensitive information

from serious decisions

information considered from the viewpoint of social fairness

ex. gender, religion, race, ethnicity, handicap, political conviction

information restricted by law or contracts

ex. insider information, customers’ private information

5 - Filter Bubble

[TED Talk by Eli Pariser]

The Filter Bubble Problem

Pariser posed a concern that personalization technologies narrow and

bias the topics of information provided to people

http://www.thefilterbubble.com/

Friend Recommendation List in Facebook

To fit for Pariser’s preference, conservative people are eliminated from

his recommendation list, while this fact is not notified to him

6 - Information Neutral Recommender System

[Kamishima+ 12]

The Filter Bubble Problem

Information Neutral Recommender System

enhancing the neutrality from a viewpoint specified by a user

and other viewpoints are not considered

fairness-aware data mining techniques

ex. A system enhances the neutrality in terms of whether

conservative or progressive, but it is allowed to make

biased recommendations in terms of other viewpoints,

for example, the birthplace or age of friends

7 - Non-Redundant Clustering

[Gondek+ 04]

non-redundant clustering : find clusters that are as independent

from a given uninteresting partition as possible

a conditional information bottleneck method,

which is a variant of an information bottleneck method

clustering facial images

Simple clustering methods find two clusters:

one contains only faces, and the other

contains faces with shoulders

Data analysts consider this clustering is

useless and uninteresting

A non-redundant clustering method derives

more useful male and female clusters,

which are independent of the above clusters

8 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

9 - Difficulty in Fairness-aware Data Mining

[Calders+ 10]

US Census Data : predict whether their income is high or low

Male

Female

High-Income

3,256 fewer

590

Low-income

7,604

4,831

Females are minority in the high-income class

# of High-Male data is 5.5 times # of High-Female data

While 30% of Male data are High income, only 11% of Females are

Occam’s Razor : Mining techniques prefer simple hypothesis

Minor patterns are frequently ignored

and thus minorities tend to be treated unfairly

10 - Red-Lining Effect

[Calders+ 10]

Calders-Verwer discrimination score (CV score)

Pr[ High-income | Male ] - Pr[ High-income | Female ]

the difference between conditional probabilities of advantageous

decisions for non-protected and protected members. The larger

score indicates the unfairer decision.

US Census Data samples

The baseline CV score is 0.19

Incomes are predicted by a naïve-Bayes classifier trained from data

containing all sensitive and non-sensitive features

The CV score increases to 0.34, indicating unfair treatments

Even if a feature, gender, is excluded in the training of a classifier

improved to 0.28, but still being unfairer than its baseline

red-lining effect : Ignoring sensitive features is ineffective against the

exclusion of their indirect influence

11 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

12 - Variables

Y

S

objective variable

sensitive feature

a binary class variable {0, 1}

a binary feature variable {0, 1}

a result of serious decision

socially sensitive information

ex., whether or not to allow credit

ex., gender or religion

X

non-sensitive features

a numerical feature vector

features other than a sensitive feature

non-sensitive, but may correlate with S

13 - Prejudice

Prejudice : the statistical dependences of an objective variable or

non-sensitive features on a sensitive feature

Direct Prejudice

Y ⊥⊥

/ S | X

a clearly unfair state that a prediction model directly depends on a

sensitive feature

implying the conditional dependence between Y and S given X

Indirect Prejudice

Y ⊥⊥

/ S | φ

statistical dependence of an objective variable on a sensitive feature

bringing red-lining effect

Latent Prejudice

X ⊥⊥

/ S | φ

statistical dependence of non-sensitive features on a sensitive feature

completely excluding sensitive information

14 - Fairness-aware Classification

True Distribution

Estimated Distribution

approximate

Pr[Y, X, S]

ˆ

Pr[Y, X, S]

=

sample

=

M[Y |X, S Pr[

]

X, S]

ˆ

M[Y |X, S Pr[

]

X, S]

learn

no indirect

no indirect

similar

prejudice

D

prejudice

Training

learn

Data

†

†

Pr [Y, X, S]

ˆ

Pr [Y, X, S]

†

=

†

=

M [Y |X, S Pr[

]

X, S]

ˆ

approximate

M [Y |X, S]Pr[X, S]

True Fair Distribution

Estimated Fair Dist.

15 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

16 - Logistic Regression

with Prejudice Remover Regularizer

modifications of a logistic regression model

add ability to adjust distribution of Y depending on the value of S

add a constraint of a no-indirect-prejudice condition

add ability to adjust distribution of Y depending on the value of S

Pr[Y =1|x, S=0] = sigmoid(x w0)

Pr[Y =1|x, S=1] = sigmoid(x w1)

weights depending on the value of S

Multiple logistic regression models are built separately, and each of

these models corresponds to each value of a sensitive feature

When predicting scores, a model is selected according to the value

of a sensitive feature

17 - No Indirect Prejudice Conditon

add a constraint of a no-indirect-prejudice condition

samples of

samples of

samples of

objective

non-sensitive

sensitive

regularization

variables

features

features

model

parameter λ

parameter

ln Pr({(y, x, s)}; ) + R({(y, x, s)}, ) +

2

2

2

η : fairness

Prejudice Remover Regularizer

parameter

L2 regularizer

the smaller value

the larger value

avoiding

more strongly constraints

more enforces

over fitting

the independence between Y and S

the fairness

18 - Prejudice Remover Regularizer

no-indirect-prejudice condition = independence between Y and S

Prejudice Remover Regularizer

mutual information between Y (objective variable) and S (sensitive feature)

ˆ

Pr[Y, S]

Pr[X, S] ˆ

M[Y |X, S] ln ˆPr[S] ˆPr[Y ]

Y

{0,1} X,S

ˆ

ˆ

Pr[y|s; ]

M[y|x, s; ] ln

ˆ

Pr[y|s; ]

Y

{0,1} (x,s)

s

expectation over X and S

is replaced with

the summation over samples

19 - Computing Mutual Information

ˆ

ˆ

Pr[y|s; ]

M[y|x, s; ] ln

ˆ

Pr[y|s; ]

Y

{0,1} (x,s)

s

This distribution can be derived by marginalizing over X

ˆ

M[y|x, s; ] ˆ

Pr[x]dx

Dom(x)

But this is computationally heavy...

approximate by the sample mean over X for each pair of y and s

ˆ

M[y

x D

|x, s; ]

Limitation: This technique is applicable only if both Y and S are discrete

20 - Calders-Verwer’s 2 Naïve Bayes

[Calders+ 10]

Unfair decisions are modeled by introducing

the dependence of X on S as well as on Y

Calders-Verwer Two

Naïve Bayes

Naïve Bayes (CV2NB)

Y

Y

S

X

S

X

S and X are conditionally

non-sensitive features X are

independent given Y

mutually conditionally

independent given Y and S

✤ It is as if two naïve Bayes classifiers are learned depending on each value of the

sensitive feature; that is why this method was named by the 2-naïve-Bayes

21 - Calders-Verwer’s Two Naïve Bayes

[Calders+ 10]

parameters are initialized by the corresponding empirical distributions

ˆ

Pr[Y, X, S] = ˆ

M[Y, S]

ˆ

Pr[X

i

i|Y, S]

M[Y, S] is modified so as to improve the fairness

estimated model ˆ

fair

M[Y, S]

fair estimated model ˆ

M†[Y, S]

keep the updated distribution similar to the empirical distribution

while CVscore > 0

if # of data classified as “1” < # of “1” samples in original data then

increase M[Y=+, S=-], decrease M[Y=-, S=-]

else

increase M[Y=-, S=+], decrease M[Y=+, S=+]

reclassify samples using updated model M[Y, S]

update the joint distribution so that its CV score decreases

22 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

23 - Experimental Conditions

Calders & Verwer’s Test Data

Adult / Census Income @ UCI Repository

Y : a class representing whether subject’s income is High or Low

S : a sensitive feature representing whether subject’s gender

X : non-sensitive features, all features are discretized, and 1-of-K

representation is used for logistic regression

# of samples : 16281

Methods

LRns : logistic regression without sensitive features

NBns : naïve Bayes without sensitive features

PR : our logistic regression with prejudice remover regularizer

CV2NB : Calders-Verwer’s two-naïve-Bayes

Other Conditions

L2 regularization parameter λ = 1

five-fold cross validation

24 - Evaluation Measure

Accuracy

How correct are predicted classes?

the ratio of correctly classified sample

NMI (normalized mutual information)

How fair are predicted classes?

mutual information between a predicted class and a sensitive

feature, and it is normalized into the range [0, 1]

I(Y ; S)

NMI =

H(Y )H(S)

25 - Experimental Results

Accuracy

Fairness (NMI between S & Y)

0.86

0.08

0.84

0.06

0.82

er

0.04

fair

0.80

NBns

LRns

high-accuracy 0.78

CV2NB

0.02

PR

0.76

0.01

0

10

20

30

0

10

20

30

fairness parameter η : the lager value more enhances the fairness

fairness parameter η

accuracy

NMI

Our method PR (Prejudice Remover) could make fairer decisions

than pure LRns (logistic regression) and NBns (naïve Bayes)

PR could make more accurate prediction than NBns or CV2NB

CV2NB achieved near-zero NMI, but PR could NOT achieve it

26 - Synthetic Data

Why did our prejudice remover fail to

make a fairer prediction than that made by CV2NB?

sensitive

si

feature

2 {0, 1} ⇠ DiscreteUniform

non-sensitive

x

independent

i = ✏i

feature

(

from si

1 + ✏

w

i, if si = 1

i =

1 + ✏

depend on si

i, otherwise

✏i ⇠ N (0, 1)

(

objective

1, if xi + wi < 0

both xi and wi

variable

yi =

0, otherwise

equally influence

an objective variable

27 - Analysis on Synthetic Data Results

M[ Y

[ | X

| , S

, =0 ]

S

M[ Y

[ | X

| , S=0 ]

S

X

W

bias

X

W

bias

η=0

11.3

11.3

-0.0257

11.3

11.4

0.0595

η=150

55.3

-53.0

-53.6

56.1

54.1

53.6

Prejudice Remover

When η=0, PR regularizer doesn’t affect weights, and these weights

for X and W are almost equal

When η=150, absolutes of weights for X are larger than those for W

PR ignores features depending on S if the influences to Y are equal

CV2NB

CV2NB treats all features equally

Prejudice Remover can consider the differences among individual

features, but CV2NB can improve fairness more drastically

28 - Outline

Applications

applications of fairness-aware data mining

Difficulty in Fairness-aware Data Mining

Calders-Verwer’s discrimination score, red-lining effect

Fairness-aware Classification

fairness-aware classification, three types of prejudices

Methods

prejudice remover regularizer, Calders-Verwer’s 2-naïve-Bayes

Experiments

experimental results on Calders&Verwer’s data and synthetic data

Related Work

privacy-preserving data mining, detection of unfair decisions,

explainability, situation testing, fairness-aware data publication

Conclusion

29 - Relation to

Privacy-Preserving Data Mining

indirect prejudice

the dependency between a objective Y and a sensitive feature S

from the information theoretic perspective,

mutual information between Y and S is non-zero

from the viewpoint of privacy-preservation,

leakage of sensitive information when an objective variable is known

differences from PPDM

introducing randomness is occasionally inappropriate for severe

decisions, such as job application

disclosure of identity isn’t problematic generally

30 - Finding Unfair Association Rules

[Pedreschi+ 08, Ruggieri+ 10]

ex: association rules extracted from German Credit Data Set

(a) city=NYC ⇒ class=bad (conf=0.25)

0.25 of NY residents are denied their credit application

(b) city=NYC & race=African ⇒ class=bad (conf=0.75)

0.75 of NY residents whose race is African are denied their credit application

conf(A ^ B ) C)

extended lift (elift)

elift =

conf(A ) C)

the ratio of the confidence of a rule with additional condition

to the confidence of a base rule

a-protection : considered as unfair if there exists association rules

whose elift is larger than a

ex: (b) isn’t a-protected if a = 2, because elift = conf(b) / conf(a) = 3

They proposed an algorithm to enumerate rules that are not a-protected

31 - Explainability

[Zliobaite+ 11]

conditional discrimination : there are non-discriminatory cases, even

if distributions of an objective variable depends on a sensitive feature

ex : admission to a university

more females apply

to medicine

explainable feature

more males apply to

sensitive feature

program

computer

gender

medicine / computer

male / female

objective variable

medicine : low acceptance

acceptance ratio

computer : high acceptance

accept / not accept

Because females tend to apply to a more competitive program,

females are more frequently rejected

Such difference is explainable and is considered as non-discriminatory

32 - Situation Testing

[Luong+ 11]

Situation Testing : When all the conditions are same other than a

sensitive condition, people in a protected group are considered as

unfairly treated if they received unfavorable decision

They proposed a method for finding

people in

unfair treatments by checking the

a protected group

statistics of decisions in k-nearest

neighbors of data points in a

protected group

Condition of situation testing is

Pr[ Y | X, S=a ] = Pr[ Y | X, S=b ] ∀ X

This implies the independence

between S and Y

k-nearest neighbors

33 - Fairness-aware Data Publishing

[Dwork+ 11]

data owner

loss function

original data

representing utilities

for the vendor

data representation

vendor (data user)

archtype

so as to maximize

vendor’s utility

under the constraints

to guarantee fairness

in analysis

fair decisions

They show the conditions that these archtypes should satisfy

This condition implies that the probability of receiving favorable

decision is irrelevant to belonging to a protected group

34 - Conclusion

Contributions

the unfairness in data mining is formalized based on independence

a prejudice remover regularizer, which enforces a classifier's

independence from sensitive information

experimental results of logistic regressions with our prejudice remover

Future Work

improve the computation of our prejudice remover regularizer

fairness-aware classification method for generative models

another type of fairness-aware mining task

Socially Responsible Mining

Methods of data exploitation that do not damage people’s lives, such

as fairness-aware data mining, PPDM, or adversarial learning,

together comprise the notion of socially responsible mining, which

it should become an important concept in the near future.

35 - Program Codes and Data Sets

Fairness-Aware Data Mining

http://www.kamishima.net/fadm

Information Neutral Recommender System

http://www.kamishima.net/inrs

Acknowledgements

We wish to thank Dr. Sicco Verwer for providing detail information

about their work and program codes

We thank the PADM2011 workshop organizers and anonymous

reviewers for their valuable comments

This work is supported by MEXT/JSPS KAKENHI Grant Number

16700157, 21500154, 22500142, 23240043, and 24500194, and JST

PRESTO 09152492

36