Classification with No Direct Discrimination

International Journal of Computer Applications Technology and Research
Volume 3– Issue 7, 464 - 467, 2014
www.ijcat.com 464
Classification with No Direct Discrimination
Deepali P. Jagtap
Department of Computer
Engineering,
KKWIEER, Nashik,
University of Pune, India
Abstract: In many automated applications, large amount of data is collected every day and it is used to learn classifier as well as to
make automated decisions. If that training data is biased towards or against certain entity, race, nationality, gender then mining model
may leads to discrimination. This paper elaborate direct discrimination prevention method. The DRP algorithm modifies the original
data set to prevent direct discrimination. Direct discrimination takes place when decisions are made based on discriminatory attributes
those are specified by the user. The performance of this system is evaluated using measures MC, GC, DDPP, DPDM etc. Different
discrimination measures can be used to discover the discrimination.
Keywords: Classifier, Discrimination, Discrimination measures, DRP algorithm, Mining model.
1. INTRODUCTION
The Latin word „Discriminare‟ is origin of the word
Discrimination, its meaning is „Distinguish between‟.
Discrimination is treating people unfairly based on their
belonging to particular group. It restrict members of certain
group from opporunities that are available to others [1].
Discrimination can also be observed in data mining. In data
mining, large amount of data is collected and is used for
training classifier. Classification model act as support to
decision making process and the basis of scoring system. This
makes business decision maker‟s work more easier [1,11]. If
that historical data itself is biased towards or against certain
entity such as gender, nationality, race, age group etc. Then
resulting mining model may show discrimination. The use of
automated decision making system gives sense of fair
decision as it does not follow any personal preference but in
actual results may be discriminatory [9]. Publishing such data
leads to discriminatory mining results. The simple solution for
discrimination prevention would consist neglecting or
removing discriminating attributes, but even after removing
those attribute, the discrimination may persist as many other
nondiscriminatory attributes may strongly co-related with
discriminatory attributes. The publicly available data may
reveal co relation between them. Also removal of sensitive
attribute results into loss of quality of original data [11].
Direct discrimination is observed when decisions
are made based on the input data containing protected groups,
whereas Indirect discrimination occurs when decisions are
made based on nondiscriminatory input data but it is strongly
or indirectly co-related with discriminatory one. For example,
If discrimination occurs against foreign worker and even after
removing that attribute from data set, one cannot guarantee
that discrimination has been prevented completely as publicly
available data may reveal nationality of that individual, hence
shows indirect discrimination.
There are various laws against discrimination, but
those are reactive not proactive. The use of technology and
new mining algorithm helps to make them proactive. Along
with mining algorithm, some algorithm and methods from
privacy preservation such as data sanitization helps to prevent
discrimination where original data is modified or support and
confidence of certain attributes is changed to make them
discrimination free. This system is useful in various
applications such as credit/insurance scoring, lending,
personnel selection and wage discrimination, job hiring, crime
detection, activities concerning public accommodation,
education, health care and many more. Benefit and services
must be made available to everyone in a nondiscriminatory
manner.
Rest of the paper is organized as follows. Section
one provides introduction, survey of literature along with pros
and cons of some of the existing methods are discussed in
section two. Section three highlights basic terminology
associated with this topic and section four describes algorithm
as well as block diagram for discrimination prevention.
Section five contains results and discussion about data set and
finally last section presents conclusion along with the future
scope of system.

Volume 3– Issue 7, 464 - 467, 2014
www.ijcat.com 465
2. LITERATURE SURVEY
Various studies have been reported the discrimination
prevention in the field of data mining. Pedreschi noticed the
discriminatory decisions in data mining based on
classification rule and discriminatory measure. The work in
this area can be traced back from year 2008. S. Ruggieri,
Pedreschi and F. Turini [14] have implemented the DCUBE.
It is oracle based tool to explore discrimination hidden in data.
Discrimination prevention can be done in three ways based on
when and in which phase data or algorithm is to be changed.
Three ways for Discrimination prevention are: Preprocessing
method, Inprocessing method and Postprocessing method.
Discrimination can be of 3 types: Direct, Indirect or
combination of both, based on presence of discriminatory
attributes and other attributes that are strongly related with
discriminatory one. Dino Pedreschi, Salvatore Ruggieri,
Franco Turini[3] has introduced a model used in Decision
Support System for the analysis and reasoning of
discrimination that helps DSS owners and control authorities
in the process of discrimination analysis [3].
Discrimination Prevention by
Preprocessing Method
In preprocessing method, the original data set is modified in
such a manner that it will not result in discriminatory
classification rule. In this method any data mining algorithm
can be applied to get mining model. Kamiran and Calder[4]
proposed a method based on “data massaging” where class
label of some of the records in the dataset is changed but as
this method is intrusive, concept of "Preferential sampling"
was introduced where distribution of objects in a given dataset
is changed to make it non-discriminatory[4]. It is based on the
idea that, "Data objects that are close to the decision boundary
are more vulnerable to be victim of discrimination." This
method uses Ranking function and there is no need to change
the class labels. This method first divides data into 4 groups
that are DP, DN, PP, PN, where first letters D and P indicate
Deprived and Privileged class respectively and second letters
P, N indicates positive and negative class label. The ranker
function then sorts data in ascending order with respect to
positive class label. Later it changes sample size in respective
group to make that data biased free. Sara Hajian and Josep
Domingo-Ferrer[9] proposed another preprocessing method to
remove direct and indirect discrimination from original
dataset. It employees 'elift' as discrimination measure to
prevent discrimination in crime and intrusion detection
system[10].
Preprocessing method is useful in applications where data
mining is to be performed by third party and data needs to
publish for public usage [9].
Inprocessing Method
Faisal Kamiran, Toon Calders and Mykola Pechenizkiy [5]
introduced algorithm based on inprocessing method using
decision Tree where instead of modifying original dataset data
mining algorithm is modified. This approach consists of two
techniques for the decision tree construction process, first is
Dependency-Aware Tree Construction and another is Leaf
Relabeling. The first technique focuses on splitting criterion
for tree construction to build a discrimination-aware decision
tree. In order to do so, it first calculates the information gain
with respect to class & sensitive attribute represented by IGC
and IGS respectively. There are three alternative criteria for
determining the best split that uses different mathematical
operation: (i) IGC-IGS (ii) IGC/IGS (iii) IGC+IGS. The
second approach consists of processing of decision tree with
discrimination-aware pruning and it relabel the tree leaves [5].
This methods requires special purpose data mining
algorithms.
Postprocessing Method
Sara Hajian, Anna Monreale, Dino Pedreschi ,Josep Domingo
Ferrer[12] proposed algorithm based on postprocessing
method that derive frequent classification rule and modifies
the final mining model using α-Protective k-Anonymous
pattern sanitization to remove discrimination from Mined
Model. Thus in postprocessing method, resultant mining
model is modified instead of modifying original data or
mining algorithm. The disadvantage of this method is, it
doesn't allow original data to be published for public usage,
and also the task of data mining should be performed by data
holder only. Toon Calders and Sicco Verwer[15] proposed
approach where the Naive Bayes classifier is modified to
perform classification that is independent with respect to a
given sensitive attribute. There are three approaches in order
to make the Naive Bayes classifier discrimination-free: (i)
modifying the probability of the decision being positive where
the probability distribution of the sensitive attribute is
modified. This method has disadvantage of either always
increasing or always decreasing the number of positive labels
assigned by the classifier, depending on how frequently the
sensitive attribute is present in dataset, (ii) training one model
for every sensitive attribute value and balancing them. This is
done by splitting the dataset into two separate sets and the
model is learned using only the tuples from the dataset that
have a favored sensitive value, (iii) adding a latent variable to
the Bayesian model. This method models the actual class
labels using a latent variable[15].
3. THEORETICAL FOUNDATION
The basic terms in data mining are described in short as
below:
3.1 Discrimination Measures
a. elift
Pedreschi [2] introduced „elift’ called extended lift as one
of the discrimination measure. For a given classification rule,
Extended lift can be calculated as below. Elift provides gain
in confidence due to presence of discriminatory item [1].
b. slift
The selection lift i.e. „slift‟ for a classification rule of
the form (A, B → C) is given as,
c. glift
Pedreschi [2] introduced „glift‟ to strengthen the notion of
α-protection. For a given classification rule glift is computed
as,

Volume 3– Issue 7, 464 - 467, 2014
www.ijcat.com 466
glift( ) = if β ≥ γ
(1 - β) / (1- γ) otherwise
3.2 Direct discrimination
Direct discrimination consists of rules or procedures that
explicitly mention disadvantaged or minority groups based on
sensitive discriminatory attributes [9]. For example, the rule r:
(Foreign_worker = Yes, City = Nasik  Hire = No) shows
direct discrimination as it contains discriminatory attribute
Foreign_worker = yes.
3.3 Indirect discrimination
Indirect discrimination consists of rules or procedures that,
while not explicitly mentioning discriminatory attributes,
intentionally or un-intentionally could generate discriminatory
decisions [9], for example, the rule r: Pin_code = 422006,
City = Nashik  Hire = No shows indirect discrimination, as
attribute Pin_code corresponds to area with mostly people
belonging to particular religion.
3.4 PD Rule
A classification rule is said to be Potentially
Discriminatory rule if it contains discriminatory item in
premise of a given rule.
3.5 PND Rule
A classification rule is said to be Potentially Non-
discriminatory rule if it doesn't contains any discriminatory
item in premise of a given rule [1, 9].
4. DISSERTATION WORK
4.1 Algorithm and Process flow
The diagram showing overall process flow of
discrimination prevention is shown in fig 1. The system takes
original dataset containing discriminatory items as an input.
4.1.1 Input Data Preprocessing
The original dataset contains numerical values so it should
be preprocessed i.e. discretization is performed on some
attributes.
Fig. 1. Flow chart for Discrimination Prevention
4.1.2 Frequent Classification Rule extraction
Later on using Apriori algorithm frequent item sets are
generated. In Apriori algorithm candidate set generation and
pruning steps are performed. The resultant frequent item sets
are used to generate frequent classification rules.
4.1.3 Discrimination Discovery Process
The frequent classification rules are then categorized into
Potentially Discriminatory and Potentially Nondiscriminatory
groups in discrimination discovery. For discrimination
discovery each classification rule is examined and is placed
into either PD or PND group based on presence of
discriminatory items in premise of the rule. If the rule found
to be PD then for every PD rule elift, glift and slift is
calculated. If that calculated value is greater than threshold
value (α) then that rule is considered as α-discriminatory.
4.1.4 Data Transformation using DRP Algorithm
Data transformation is carried out in the next step as α-
discriminatory rules need to be treated further to remove
discrimination where class label of some of the records is
perturbed to prevent discrimination. As a result of above
process finally the transformed dataset is obtained as an
output.
Data transformation is second step in discrimination
prevention where the data is actually modified to make it
biased free. In this step modifications are done using the
definition of elift/glift/slift i.e. equality constraint on rule are
enforced to satisfy the definition of corresponding
discrimination prevention measure.

Volume 3– Issue 7, 464 - 467, 2014
www.ijcat.com 467
Direct Rule Protection (DRP) algorithm is used here that
converts α-discriminatory rules into α-protective rule using
the definition of elift. It can be done in following way:
let r': α-discriminatory rule, condition enforced on r' is:
= elift(r') < α
=
= confidence(r': A, B→C)/confidence(B→C) <α
= confidence(r': A, B→C/α <confidence (B→ C)
Here one needs to increase confidence (B→C), so change
the class item from ¬C to C for all records in original DB that
supports the rule of the form(¬A,B → ¬C). In this way this
method changes the class label of class item in some
records[9]. Similar method for slift as well as glift can be
carried out.
4.2 Performance measures
To measure the success of the method in removing all
evidence of Direct Discrimination and to measure quality of
the modified data, following measures are used:
4.2.1 Direct discrimination prevention degree
(DDPD)
The DDPD counts the percentage of α-discriminatory rules
that are no longer α-discriminatory in the transformed data set.
4.2.2 Direct discrimination protection preservation
(DDPP)
This measure counts the percentage of the α-protective rules
in the original data set that remain α-protective in the
transformed data set.
4.2.3 Misses cost (MC)
This measure helps to find the percentage of rules that are
extractable from the original data set but cannot be extracted
from the transformed data set. This is considered as side effect
of the transformation process.
4.2.4 Ghost cost (GC)
This ghost cost quantifies the percentage of the rules that
are extractable from the transformed data set but were not
extractable from the original data set.
This MC and GC are the measures that are used in the
context of privacy preservation. As similar approach of data
sanitization is used in some methods for discrimination
prevention, the same measures that are MC and GC can be
applied to find out the information loss [16].
5. RESULTS AND DISCUSSION
German Credit Data set
This data set consists of 1000 records as well as 20
attributes. Out of those 20 attributes 7 are numerical and
remaining 13 are categorical attributes. The class attributes
indicates good or bad class for given bank account holder.
Here the attribute foreign worker = Yes, Personal status =
Female but not single and age = old are considered as
discriminatory items where age > 50 is considered as old.
The Table I show the partial results computed on German
credit dataset containing total number of classification rules
generated and number of α-Discriminatory rules and the
number of lines modified in original data set.
Table 1. German Credit dataset: Columns show the
partial results for No. of α-Discriminatory rules, No. of
lines modified.
Total No. of
Classification
rules
No. of α-
Discriminatory
rules
No. of Lines
modified
9067 45 49
6. CONCLUSION
It is very important to remove discrimination, which can be
observed in data mining, from original data. The removal of
discriminatory attributes does not solve the problem. In order
to prevent such discrimination, Discrimination Prevention by
preprocessing technique is advantageous over the other two
methods. The approach mentioned in this paper works in two
steps: first is the discrimination discovery where α-
Discriminatory rules are extracted and then in second step
data transformation is performed in which the original data is
transformed to prevent direct discrimination. This second step
follows similar approach of Data Sanitization that is used in
privacy preservation context. Many such algorithms uses 'elift'
as a measure of discrimination, but instead of that one may
use slift, glift as a measure of discrimination. The
performance measure metrics i.e. DDPD, DDPP, MC, GC
analyses data to check quality of transformed data as well as
presence of direct discrimination. The less number of
classification rules will be extracted from transformed data set
as compare to original data set. The use of different
discrimination measures such as slift, glift results into varying
number of discriminatory rules and it have varying impact on
original data.
In the future, one may explore how rule hiding in privacy
preservation or other privacy preserving algorithms helps to
prevent discrimination.
7. ACKNOWLEDGMENTS
With deep sense of gratitude I thank to my guide
Prof. Dr. S. S. Sane, Head of Department of Computer
Engineering, (KKWIEER, Nashik, University of Pune, India)
for guiding me and his constant support and valuable
suggestions that have helped me in the successful completion
of this paper.
8. REFERENCES
[1] D. Pedreschi, S. Ruggieri, and F. Turini, "Discrimination-
Aware Data Mining," Proc. 14th ACM Int'l Conf. Knowledge
Discovery and Data Mining (KDD'08), pp. 560-568, 2008.
(Cited by 56)
[2] D. Pedreschi, S. Ruggieri, and F. Turini, “Measuring
Discrimination in Socially-Sensitive Decision Records,” Proc.
Ninth SIAM Data Mining Conf. (SDM ‟09),pp. 581-592,
2009.
[3] D. Pedreschi, S. Ruggieri, and F. Turini, “Integrating
Induction and Deduction for Finding Evidence of
Discrimination,”Proc. 12thACM Int‟l Conf. Artificial
Intelligence and Law (ICAIL ‟09), pp. 157-166, 2009.
[4] F. Kamiran and T. Calders, "Classification with no
Discrimination by Preferential Sampling,"Proc. 19th Machine
Learning Conf. Belgium and TheNetherlands, 2010.
[5] F. Kamiran, T. Calders, and M. Pechenizkiy,
"Discrimination Aware Decision Tree Learning,"Proc. IEEE
Int'l Conf. Data Mining (ICDM '10), pp. 869-874, 2010.

Volume 3– Issue 7, 464 - 467, 2014
www.ijcat.com 468
[6] M.Kantarcioglu, J. Jin and C. Clifton. When do data
mining results violate privacy? In KDD 2004, pp. 599-604.
ACM, 2004.
[7] P. N. Tan, M. Steinbach and V. Kumar, “Introduction to
Data Mining”.Addison-Wesley, 2006.
[8] R. Agrawal and R. Srikant, "Fast Algorithms for Mining
Association Rules in Large Databases,"Proc. 20th Int'l Conf.
Very Large Data Bases, pp. 487-499, 1994.
[9] S. Hajian and J. Domingo, "A Methodology for Direct and
Indirect Discrimination prevention in data mining." IEEE
transaction on knowledge and data engineeing, VOL. 25, NO.
7, pp. 1445-1459, JULY 2013. (Cited by 12)
[10] S. Hajian, J. Domingo-Ferrer, and A. Martnez Balleste,
“Discrimination Prevention in Data Mining for Intrusion and
Crime Detection,"Proc. IEEE Symp. Computational
Intelligence in Cyber Security (CICS '11), pp. 47-54, 2011.
[11] S. Hajian, J. Domingo-Ferrer, and A. Martı´nez-
Balleste´, “Rule Protection for Indirect Discrimination
Prevention in Data Mining,”Proc. Eighth Int‟l Conf. Modeling
Decisions for Artificial, Intelligence (MDAI ‟11),pp. 211-222,
2011,
[12] S. Hajian, A. Monreale, D. Pedreschi, J. Domingo-Ferrer
and F. Giannotti. “Injecting discrimination and privacy
awareness into pattern discovery,” In 2012 IEEE 12th Inter-
national Conference on Data Mining Workshops, pp. 360-369.
IEEE Computer Society, 2012.
[13] S. Ruggieri, D. Pedreschi and F. Turini. “Data mining for
discrimination discovery,” ACM Transactions on Knowledge
Discovery from Data (TKDD), 4(2), Article 9, 2010.
[14] S. Ruggieri, D. Pedreschi, and F. Turini, "DCUBE:
Discrimination Discovery in Databases,"Proc. ACM Int‟l
Conf. Management of Data (SIGMOD‟10), pp. 1127-1130,
2010.
[15] T. Calders and S. Verwer. “Three naive Bayes
approaches for discrimination-free classification,” Data
Mining and Knowledge Discovery, 21(2):277-292, 2010.
(Cited by 44)
[16] V. Verykios and A. Gkoulalas Divanis, "A Survey of
Association Rule Hiding Methods for Privacy," Privacy-
Preserving Data Mining: Models and Algorithms,C.C.
Aggarwal and P.S. Yu, eds.,Springer, 2008.

Classification with No Direct Discrimination

More Related Content

What's hot (17)

Viewers also liked (9)

Similar to Classification with No Direct Discrimination (20)

More from Editor IJCATR (20)

Recently uploaded (20)

Classification with No Direct Discrimination