Content uploaded by Roberto Saia
Author content
All content in this area was uploaded by Roberto Saia on Jul 26, 2017
Content may be subject to copyright.
Evaluating Credit Card Transactions in the Frequency Domain
for a Proactive Fraud Detection Approach
Roberto Saia and Salvatore Carta
Dipartimento di Matematica e Informatica, Universit`a di Cagliari, Italy
{roberto.saia,salvatore}@unica.it
Keywords: Business Intelligence, Fraud Detection, Pattern Mining, Fourier, Metrics.
Abstract: The massive increase in financial transactions made in the e-commerce field has led to an equally massive
increase in the risks related to fraudulent activities. It is a problem directly correlated with the use of credit
cards, considering that almost all the operators that offer goods or services in the e-commerce space allow
their customers to use them for making payments. The main disadvantage of these powerful methods of
payment concerns the fact that they can be used not only by the legitimate users (cardholders) but also by
fraudsters. Literature reports a considerable number of techniques designed to face this problem, although
their effectiveness is jeopardized by a series of common problems, such as the imbalanced distribution and the
heterogeneity of the involved data. The approach presented in this paper takes advantage of a novel evaluation
criterion based on the analysis, in the frequency domain, of the spectral pattern of the data. Such strategy
allows us to obtain a more stable model for representing information, with respect to the canonical ones,
reducing both the problems of imbalance and heterogeneity of data. Experiments show that the performance
of the proposed approach is comparable to that of its state-of-the-art competitor, although the model definition
does not use any fraudulent previous case, adopting a proactive strategy able to contrast the cold-start issue.
1 INTRODUCTION
The credit card frauds (i.e., when someone makes
purchases without authorization or counterfeits a
credit card) represent one of the major issues that af-
fect the e-commerce environment, especially in this
age of exponential growth of it. Authoritative studies
performed by American Association of Fraud Exam-
iners1report that this kind of fraud involves the 10-
15% of all fraud cases, for a total financial value close
to 75-80%. Only in the United States of America, this
scenario leads toward an estimated average loss per
fraud case of 2million of dollars.
Such situation has given rise to a great interest by
the research community in order to develop increas-
ingly effective techniques able to detect the fraudu-
lent transactions. This is a task that can be performed
by exploiting several techniques but there are some
common issues that the researchers must face. The
most important between them are the imbalanced dis-
tribution and the heterogeneity of the involved data.
The first issue is given by the fact that the fraudu-
lent transactions are usually less than the legitimate
1http://www.acfe.com
ones, configuring an unbalanced distribution of data
that reduces the effectiveness of the machine learning
strategies (Japkowicz and Stephen, 2002).
The common criterion used in almost all the state-
of-the-art approaches of fraud detection is substan-
tially based on the comparison between the set of pre-
vious legitimate transactions of a user and the new
transactions under evaluation. This is a rather trivial
criterion that in many cases, due to the high hetero-
geneity of data, leads toward misclassifications. In
order to overcome this problem, a fraud detection ap-
proach should be able to use as much as possible in-
formation about the transactions during the evaluation
process, but this is not always possible due to the in-
ability of some approaches to manage some informa-
tion (e.g., Random Forests, one of the most perform-
ing approaches, is not able to manage types of data
that involve a large number of categories).
The idea around which this paper was born
is the introduction of a new model of evalua-
tion based on the spectral pattern of the trans-
actions, a novel representation of the informa-
tion obtained by exploiting the Fourier transforma-
tion (Duhamel and Vetterli, 1990). This operation of-
fers us a new point of view on data, providing the
following advantages: considering that our process
involves only the previous legitimate transactions, it
allows us to operate proactively; such proactivity ef-
fectively face the cold-start issue (i.e., scarcity or ab-
sence of fraudulent cases during the model training);
the representation in the frequency domain reduces
the issues related to the data heterogeneity, since the
spectral model is less influenced by the data varia-
tions.
The main contributions of this paper are summa-
rized below:
• definition of the time series to use in a Fourier
Transform process, in terms of sequence of values
assumed by the transaction features;
• formalization of the comparison process between
the spectral pattern of an unevaluated transaction
and those of the previous legitimate transactions;
• definition of an algorithm able to to classify a new
transaction as legitimate or fraudulent, on the ba-
sis of the previous comparison process.
The rest of the paper is organized as follows: Sec-
tion 2 presents the background and related work of the
scenario taken into account; Section 3 provides a for-
mal notation, makes some assumptions, and defines
the faced problem; Section 4 describes the proposed
approach; Section 5 provides details on the experi-
mental environment, on the used datasets and met-
rics, as well as on the adopted strategy and selected
competitor, presenting the experimental results; some
concluding remarks and future work are given in the
last Section 6.
2 BACKGROUND AND RELATED
WORK
The main task of a fraud detection system is the
evaluation of a new financial transaction with the aim
to classify it as legitimate or fraudulent, by using
the information gathered in the past (i.e. value of
the features that compose each transaction and if it
was a fraud or not). This section provides a general
overview of the context taken into account in this pa-
per, starting with the introduction of the most used
strategies and approaches, continuing with the de-
scription of the outstanding problems, and conclud-
ing with the presentation of core concepts on which
the proposed approach is based, giving some details
about the state-of-the-art approach chosen to evaluate
its performance.
2.1 Strategies and Approaches
Operative Strategies. The fraud detection ap-
proaches operate by following a supervised or unsu-
pervised strategy (Phua et al., 2010).
Asupervised strategy works by exploiting the
previous fraudulent and non-fraudulent transactions
gathered by the system, using them in order to define
a model able to classify the new transactions in a spe-
cific class (i.e., legitimate or fraudulent). It is evident
how such strategy needs a series of examples concern-
ing both classes, and how its effectiveness is limited
to the recognition of known patterns.
An unsupervised strategy operates instead by ana-
lyzing the new transactions in order to evaluate when
they present anomalies in their values, compared to
the typical range of values that characterizes the con-
text taken into account. It is an inefficient strategy
because a fraudster can operate in order to avoid that
the transaction presents anomalies in its values, and
for this reason the definition of effective unsupervised
strategies represents a hard challenge.
Operative Approaches. The most common way to
operate in order to detect fraudulent events in a finan-
cial data stream related to a credit card activity is the
adoption of a static approach (Pozzolo et al., 2014).
It operates by dividing the data stream into blocks of
equal size, training its model by using only a limited
number of initial and contiguous blocks. A differ-
ent modality is instead adopted by the updating ap-
proach (Wang et al., 2003), where at each new block,
the model is updated by training it by using a certain
number of latest and contiguous blocks. The forget-
ting approach (Gao et al., 2007) represents another
possible operative way. By following this approach,
the user model is updated when a new block appears,
by using only the legitimate transactions present in the
last two blocks, but by using all the previous fraudu-
lent transactions. The models generated by these ap-
proaches can be directly exploited in order to evaluate
the future blocks, or they can be used to define a big-
ger model of evaluation.
The main problems related to the aforementioned
approaches are the following: the static approach is
not able to model the users behavior; the updating
approach is not able to operate with small classes of
data; the forgetting approach presents a high com-
putational complexity. In addition, these approaches
have to face the common issues described in the next
Section 2.2.
2.2 Open Problems
Data Scarcity Issue. The task of the researchers
working in this area is complicated by the scarcity of
public real-world datasets. This mainly happens due
to the restrictive policies adopted by those working in
this field, which do not allow them to release informa-
tion about their business activities for privacy, com-
petition, or legal issues. Not even a release in anony-
mous form of the data is usually taken into account
by many financial operators, since also in anony-
mous form such data can provide precious informa-
tion about their customers, and thus they could reveal
potential vulnerabilities of the related e-commerce in-
frastructure.
Non-adaptability Issue. This problem concerns the
inability of the fraud detection models to correctly
classify the new transactions, when their features give
rise to different patterns (wrt the patterns used to
define the evaluation model). Both the supervised
and unsupervised fraud detection approaches are
affected by this problem (Sorournejad et al., 2016),
which leads toward misclassifications, due to their in-
ability to detect new legitimate or fraudulent patterns.
Data Heterogeneity Issue. Pattern recognition rep-
resents a very important branch of the machine learn-
ing, since it can be used to solve a large number of
real-world problems. The effectiveness of these pro-
cesses is nevertheless jeopardized by the heterogene-
ity of the involved data. Such problem happens due
to incompatibility between similar features resulting
in the same data being represented differently in dif-
ferent datasets (Chatterjee and Segev, 1991).
Data Unbalance Issue. Another important issue that
the approaches of fraud detection have to face is the
unbalanced distribution of data during the training of
their evaluation models. This means that the informa-
tion available to train an evaluation model are typi-
cally composed by a large number of legitimate cases
and a small number of fraudulent ones, a data con-
figuration that reduces the effectiveness of the clas-
sification approaches (Japkowicz and Stephen, 2002;
Brown and Mues, 2012). A common strategy
adopted in order to face this problem is the artificial
balance of data (Vinciotti and Hand, 2003), made by
performing an over-sampling or under-sampling pro-
cess: in the first case the balance is obtained by dupli-
cating some of the transactions that are less in num-
ber (usually, the fraudulent ones), while in the second
case it is obtained by removing some of the transac-
tions that are in greater number (usually, the legiti-
mate ones).
Cold-start Issue. The cold-start problem arises when
the set of data used to train an evaluation model
does not contain enough information about the do-
main taken into account, making it impossible to de-
fine a reliable model (Donmez et al., 2007). In other
words, it happens when the training data are not rep-
resentative of all the involved classes of informa-
tion (Attenberg and Provost, 2010) (i.e., in our case,
legitimate and fraudulent). The approach presented in
this paper faces this problem by training its evaluation
model by using only a class of data (i.e., the legitimate
cases). It represents a side effect of the adopted proac-
tive approach that allows a system to operate without
the need to have previous fraudulent transactions as
examples, with all the advantages that derive from it.
2.3 Proposed Approach and Competitor
Frequency Spectrum Evaluation. The idea that
composes the core of this work is to perform the eval-
uation process in the frequency domain, by defining
the evaluation model in terms of frequency compo-
nents (spectral pattern). Such operation is performed
by considering the sequence of values assumed by the
transaction features as a time series, moving its ana-
lysis from the canonical domain to the frequency one.
In order to perform this operation we use the Dis-
crete Fourier Transform (DFT ), whose formalization
is shown in Equation 1, where iis the imaginary unit.
Fn
def
=∑N−1
k=0fk·e−2πink/N,n∈Z(1)
fk=1
N∑N−1
n=0Fn·e2πikn/N,n∈Z(2)
The result is a set of sinusoidal functions, as
shown in Figure 1, each of them related to a specific
frequency component. We can return to the original
time domain by using the inverse Fourier transform
reported in Equation 2. In the context of the presented
approach we use the Fast Fourier Transform (FFT )
algorithm in order to perform the Fourier transforma-
tions, since it allows us to rapidly compute the DFT
by factorizing the input matrix into a product of sparse
(mostly zero) factors. This is a largely used algorithm
because it is able to reduce the computational com-
plexity of the process from O(n2)to O(nlog n), where
ndenotes the data size.
Random Forests Approach. The Random
Forests (Breiman, 2001), approach represents one of
the most common and powerful state-of-the-art tech-
niques for data analysis, because in most of the cases
it outperforms the other ones (Lessmann et al., 2015;
Brown and Mues, 2012; Bhattacharyya et al., 2011).
It consists in an ensemble learning method for clas-
sification and regression based on the construction
of a number of randomized decision trees during
the training phase. The conclusion are inferred by
averaging the obtained results and this technique can
f
t
m
f1
f2
...
fX
Figure 1: T ime and F requency Domains
be used to solve a wide range of prediction problems,
with the advantage that it does not need any complex
configuration.
3 PRELIMINARIES
This section provides the formal notation adopted
in this paper and some basic assumptions, as well as
the definition of the faced problem.
3.1 Formal Notation
Given a set of classified transactions T=
{t1,t2,...,tN}, and a set of features V={v1,v2,
.. . , vM}that compose each t∈T, we denote as
T+={t1,t2,...,tK}the subset of legitimate transac-
tions (then T+⊆T), and as T−={t1,t2,...,tJ}the
subset of fraudulent ones (then T−⊆T). We also
denote as ˆ
T={ˆ
t1,ˆ
t2,...,ˆ
tU}a set of unclassified
transactions. It should be observed that a trans-
action only can belong to one class c∈C, where
C={legitimat e,f raud ulent}. Finally, we denote
as F={f1,f2,..., fX}the frequency components
(spectrum) obtained as result of the DF T process.
3.2 Assumptions
A periodic wave is characterized by a frequency f
and a wavelength λ(i.e., the distance in the medium
between the beginning and end of a cycle λ=w
f0,
where wstands for the wave velocity), which are
defined by the repeating pattern. The non-periodic
waves that we take into account during the Discrete
Fourier Transform process do not have a frequency
and a wavelength. Their fundamental period τis the
period where the wave values were taken and sr de-
notes their number over this time (i.e., the acquisition
frequency).
Assuming that the time interval between the ac-
quisitions is equal, on the basis of the previous defi-
nitions applied in the context of this paper, the con-
sidered non-periodic wave is given by the sequence
of values assumed by each distinct feature v∈Vthat
characterize the transactions in the set T+(i.e., the
past legitimate transactions), and this sequence of val-
ues represents the time series taken into account in the
DFT process. Their fundamental period τstarts with
the value assumed by the feature in the oldest transac-
tion of the set T+and it ends with the value assumed
by the feature in the newest transaction, thus we have
that sr =|T+|; the sample interval si is instead given
by the fundamental period τdivided by the number
of acquisition, i.e., si =τ
|T+|. The frequency-domain
representation, obtained by the DFT, process gives us
information about the magnitude and phase of the sig-
nal at each frequency.
Denoting as xthe output of the process, it repre-
sents a series of complex numbers, where xris the
real part and xiis the imaginary one (i.e., we have that
x= (xr+ixi)). The magnitude can be calculated by
using |x|=q(x2
r+x2
i)and that the phase can be cal-
culated by using ϕ(x) = arctan xi
xr. In our approach
we take into account only the frequency magnitude.
3.3 Problem Definition
Denoting as Ξthe process of comparison between
all the spectral patterns related to the time series
extracted from the set T+and all the spectral pat-
terns related to the time series extracted from the set
of unevaluated transactions ˆ
T(taken one at a time),
our approach is aimed to classify as legitimate or
fraudulent each transaction ˆ
t∈ˆ
T. Given a function
EVAL(ˆ
t,Ξ)able to verify the classification correct-
ness of a transaction ˆ
t∈ˆ
Tmade by using our ap-
proach, which returns a boolean value β(0=misclas-
sification,1=correct classification), our objective can
be formalized in terms of maximization of the results
sum, as shown in Equation 3.
max
0≤β≤| ˆ
T|
β=
|ˆ
T|
∑
u=1
EVAL(ˆ
tu,Ξ)(3)
4 PROPOSED APPROACH
The implementation of the proposed approach was
performed through the following three steps, which
will be explained later:
1. Data Definition: definition of the time series in
terms of sequence of values assumed by the trans-
action features;
2. Data Processing: conversion of the time series in
the frequency spectrum by recurring to the DFT
process;
3. Data Evaluation: formalization of the algorithm
able to classify a new transaction as legitimate or
fraudulent on the basis of a spectrum comparison
process.
4.1 Data Definition
Formally, a time series is a series of data points stored
by following the time order and, in most of the cases,
it is a sequence of discrete-time data measured at suc-
cessive equally spaced points in time. In the context
of our approach, we considered as time series (ts) the
sequence of values v∈Vassumed by the features of
the transactions in T+and ˆ
T, as shown in Equation 4.
The time series related to an item ˆ
t∈ˆ
Twill be com-
pared to the time series related to all the items t+∈T+,
by following the criteria explained in the next steps.
T+=
v1,1v1,2... v1,M
v2,1v2,2... v2,M
.
.
..
.
.....
.
.
vK,1vK,2... vK,M
ˆ
T=
v1,1v1,2... v1,M
v2,1v2,2... v2,M
.
.
..
.
.....
.
.
vU,1vU,2... vU,M
ts(T+) = (v1,1,v2,1,...,vK,1),(v1,2,v2,2,...,vK,2),···,(v1,M,v2,M,...,vK,M)
ts(ˆ
T) = (v1,1,v2,1,...,vU,1),(v1,2,v2,2,...,vU,2),···,(v1,M,v2,M,...,vU,M)
(4)
4.2 Data Processing
In this step, we move the time series of the trans-
actions to the frequency domain by a DFT process
performed through the F FT approach introduced in
Section 5. In a preliminary study we compared the
transaction representations in the time domain (time
series) to those in the frequency domain (frequency
spectrum). Without going deep into the merits of the
formal characteristics of a Fourier transformation but
by limiting our analysis at the context taken into ac-
count, in our approach we want to exploit the follow-
ing two properties:
1. Phase invariance: the first property demonstrates
that there are not variations in the spectral pattern
in case of value translation2. More formally, it is
one of the phase properties of the Fourier trans-
form, i.e., a shift of a time series in the time do-
main leaves the magnitude unchanged in the fre-
quency domain (Smith et al., 1997). It means that
the representation in the frequency domain allows
us to detect a specific pattern, regardless the posi-
tion of the values assumed by the transaction fea-
tures that originate it;
2. Amplitude correlation: the second property in-
stead proves the existence of a direct correla-
tion between the values assumed by the fea-
tures in the time domain and the corresponding
2A translation in time domain corresponds to a change
in phase in the frequency domain.
magnitudes assumed by the spectral components
in the frequency domain. More formally, it is
the homogeneity property of the Fourier trans-
form (Smith et al., 1997), i.e., when the amplitude
is altered in one domain, it is altered by the same
entity in the other domain3. This ensures that the
proposed approach is able to differentiate identi-
cal spectral patterns on the basis of the values as-
sumed by their transaction features.
4.3 Data Evaluation
The evaluation of a new transaction ˆ
t∈ˆ
Tis performed
by comparing its spectral pattern F(ˆ
t)(i.e., the series
of values f∈F) to that of each previous legitimate
transactions t+∈T+. This is done by using the cosine
similarity metric described in Section 5, as shown in
Equation 5, where ∆represents the similarity value in
terms of cosine similarity,αa threshold value exper-
imentally defined in Section 5, and cis the resulted
classification.
∆=cos(F(t),F(ˆ
t)),with c =(∆≥α,legitimate
∆<α,fraudulent (5)
The final classification of a new transaction ˆ
t,
which take into account all the comparisons (Equa-
tion 5) between the transaction ˆ
tand all the transac-
tions in T+, is performed by using the Algorithm 1.
It takes as input the set T+of past legitimate trans-
actions, a transaction ˆ
tto evaluate, and the thresh-
old value αto use in the spectral pattern compar-
ison process. It returns a boolean value that in-
dicates the ˆ
tclassification (i.e., true=legitimate or
false=fraudulent). It should be noted that, if it needs,
the time complexity of the algorithm can be reduced
by parallelizing the process over several machines
(e.g., by exploiting large scale distributed computing
models).
5 EXPERIMENTS
This section reports information about the exper-
imental environment, the used datasets and metrics,
the adopted strategy, and the results of the performed
experiments.
5.1 Environment
The proposed approach was developed in Java, where
we use the JTransforms4library to operate the Fourier
3Scaling in one domain corresponds to scaling in the
other domain
4https://sourceforge.net/projects/jtransforms/
Algorithm 1 Transaction evaluation
Input: T+=Legitimate previous transactions, ˆ
t=Unevaluated transaction,
α=Threshold value
Output: β=Classification of the transaction ˆ
t
1: procedure TRANS ACT ION EVALUAT ION (T+,ˆ
t)
2: ts1←getT imeseries(ˆ
t)
3: sp1←getS pectrum(ts1)
4: for each t+in T+do
5: ts2←getT imeseries(t+)
6: sp2←getS pectrum(ts2)
7: cos ←cos +getCosineSimilarit y(sp1,s p2)
8: end for
9: avg ←cos
|T+|
10: if avg >αthen
11: β←true
12: else
13: β←f alse
14: end if
15: return β
16: end procedure
transformations. The state-of-the-art approach (i.e.,
Random Forests) and the metrics to evaluate it have
been implemented in R5, by using randomForest,
DMwR, and ROCR packages. The RF parameters
have been tuned by searching those that maximize the
performance. For reasons of reproducibility of the RF
experiments, the Rfunction set.seed() has been used
in the code to fix the seed of the random number gen-
erator.
5.2 DataSet
The real-world dataset used for the evaluation of the
proposed approach is related to a series of credit
card transactions made by European cardholders6. In
more detail, this dataset contains the transactions car-
ried out in two days of September 2013, for a to-
tal of 492 frauds out of 284,807 transactions. It
should be observed how it represents an highly unbal-
anced dataset (Pozzolo et al., 2015), considering that
the fraudulent cases are only the 0.0017% of all the
transactions.
5.3 Metrics
Cosine Similarity. The cosine similarity (Cosim) be-
tween two non-zero vectors ~v1and ~v2is calculated in
terms of cosine angle between them, as shown in the
Equation (6). It allows us to evaluate the similarity
between two spectral patterns by comparing the vec-
tors given by the magnitude of their frequency com-
ponents.
5https://www.r-project.org/
6https://www.kaggle.com/dalpozz/creditcardfraud
Cosim(~v1, ~v2) = cos(~v1, ~v2) = ~v1·~v2
k~v1k·k~v2k(6)
Accuracy. The Accuracy metric reports the number
of transactions correctly classified, compared to the
total number of them. Given a set of transactions ˆ
Tto
be classified, it is calculated as shown in Equation 7,
where |ˆ
T|stands for the total number of transactions,
and |ˆ
T(+)|for the number of those correctly classified.
Accuracy(ˆ
T) = |ˆ
T(+)|
|ˆ
T|(7)
Sensitivity. The Sensitivity metric measures the num-
ber of transactions correctly classified as legitimate,
providing an important information, since it allows us
to evaluate the predictive power of our approach in
terms of capability to identify the legitimate transac-
tions. Given a set of transactions ˆ
Tto be classified,
the Sensitivity is calculated as shown in Equation 8,
where |ˆ
T(T P)|stands for the number of transactions
correctly classified as legitimate and |ˆ
T(FN )|for the
number of legitimate transactions wrongly classified
as fraudulent.
Sensitivit y(ˆ
T) = |ˆ
T(TP)|
|ˆ
T(TP)|+|ˆ
T(FN )|(8)
F-score. The F-score is considered an ef-
fective performance measures for unbalanced
datasets (Pozzolo et al., 2015). It represents the
weighted average of the Precision and Recall metrics
and it is a largely used metric in the statistical analysis
of binary classification, returning a value in a range
[0,1], where 0 is the worst value and 1 the best one.
Given two sets T(P)and T(R), where T(P)denotes the
set of performed classifications of transactions, and
T(R)the set that contains the actual classifications of
them, this metric is defined as shown in Equation 9.
F-score(T(P),T(R)) = 2·Precision·Recall
Precision+Recal
with
Precision(T(P),T(R)) = |T(R)∩T(P)|
|T(P)|
Recall(T(P),T(R)) = |T(R)∩T(P)|
|T(R)|
(9)
AUC. The Area Under the Receiver Operating Char-
acteristic curve (AUC) is a performance measure
used to evaluate the effectiveness of a classification
model (Faraggi and Reiser, 2002). Its result is in a
range [0,1], where 1 indicates the best performance.
Given the subset of previous legitimate transactions
T+and the subset of previous fraudulent ones T−, the
formalization of the AUC metric is reported in the
Equation 10, where Θindicates all possible compar-
isons between the transactions of the two subsets T+
and T−. The result is obtained by averaging over these
comparisons.
Θ(t+,t−) =
1,i f t+>t−
0.5,i f t+=t−
0,i f t+<t−
AUC =1
|T+|·|T−|
|T+|
∑
1
|T−|
∑
1
Θ(t+,t−)(10)
5.4 Strategy
Cross-validation. In order to reduce the impact of
data dependency, improving the reliability of the ob-
tained results, all the experiments have been per-
formed by using the k-fold cross-validation criterion,
with k=10. Each dataset is divided in ksubsets, and
each ksubset is used as test set, while the other k-
1subsets are used as training set. The final result is
given by the average of all results.
Threshold Tuning. Before starting the experiments
we need to identify the optimal threshold αto use in
the evaluation process, according to the Equation 5.
In order to maintain a proactive approach, we perform
this operation by using only the previous legitimate
transactions, calculating the average value of the co-
sine similarity related to all pairs of different trans-
actions t+∈T+.Through this process we want to ob-
tain the average value of cosine similarity measured
between the frequency-domain representation (spec-
tral pattern) of all pairs of previous legitimate trans-
actions, so that we can use it to identify the potential
fraudulent transactions. The results indicate 0.901 as
optimal threshold (the recalculation frequency of that
value depends on the context in which we operate).
5.5 Competitor
We compare our approach to Random Forests one. It
is implemented in Rlanguage, by using the random-
Forest and DMwR packages. The DMwR package
allows Random Forests to manage the class imbal-
ance problem through the Synthetic Minority Over-
sampling Technique (SMOTE) (Bowyer et al., 2011).
It represents a very popular sampling technique able
to create new synthetic data by randomly interpolat-
ing pairs of nearest neighbors. The combined use of
Random Forests and SMOTE allows us to verify the
performance of our approach compared to one of the
best solutions for fraud detection at the state of the art.
5.6 Results
The first set of experiments, the results of which are
shown in Figure 2, was aimed to evaluate the capabil-
ity of the spectral pattern to model a class of trans-
actions (in our case, the legitimate one), compared
to that of a canonic approach. We compared a mil-
lion of transaction pairs (t′
+,t′′
+)(with t′
+,t′′
+∈T+and
0 0.2 0.40.60.8 1
·106
0.70
0.80
0.90
1.00
Transact ions
αvalue
Time domain Frequency domain
Figure 2: Model E valuation
t′
+6=t′′
+), measuring the cosine similarity of the rel-
ative time series and spectral patterns. We observe
how, since the beginning, the spectral pattern is able
to effectively represent the class of legitimate trans-
actions, since the variation of the cosine similarity is
much smaller than that measured by comparing the
same time series, providing a more stable modeliza-
tion of the transaction class.
The second set of experiments was instead fo-
cused on the evaluation of the proposed approach in
terms of Accuracy,Sensitivity, and F-score. The re-
sults show in Figure 3 indicate that the FD perfor-
mance are similar to those of Random Forests, al-
though it does not use any previous fraudulent trans-
action to train its model, adopting a pure proactive
strategy. It means that our approach is able to operate
without training its model with both classes of trans-
actions (legitimate and fraudulent).
The last set of experiments was aimed to evalu-
ate the performance of the FD approach in terms of
AUC. This metric is aimed to evaluate the predic-
tive power of a classification model and the results in
Figure 3 show how our model achieves performance
close to those of RF, also considering that we do
not exploit previous fraudulent transactions during the
model training.
Summarizing, we can observe how the spec-
tral representation of the transactions faces the non-
adaptability and heterogeneity issues, thanks to its
stability. We can also observe how the proactive
strategy followed by our approach is able to re-
duce/overcome the data imbalance and cold-start is-
sues, since only a class of transactions is used. Such
proactivity allows a real-world fraud detection system
to operate even in the absence of previous cases of
fraudulent cases, with all the obvious advantages that
derive from it.
0.20 0.40 0.60 0.80 1.00
RF
FD
0.95
0.77
0.91
0.76
0.95
0.87
0.98
0.77
Value
Ap proaches
Accuracy Sensitivity F-score AUC
Figure 3: Per f ormance
6 CONCLUSIONS AND FUTURE
WORK
Credit card fraud detection systems play a cru-
cial role in our e-commerce age, where an increasing
number of transactions takes place through this pow-
erful instrument of payment, with all the risks that
it involves. More than wanting to replace the exist-
ing state-of-the-art solutions, the approach presented
in this paper wants to introduce a novel frequency-
domain-based model that allows a fraud detection
system to operate proactively. The results obtained
are interesting, since it is necessary to consider that
the state-of-the-art competitor taken into account (i.e.,
Random Forests), in addition to using both classes of
transactions to train its model also preprocesses the
dataset by using a balancing technique (i.e., SMOTE).
It should be noted that the credit card context taken
into account is only one of the possible scenarios,
since the proposed approach can be used in any con-
text characterized by financial electronic transactions.
A possible future work could take into account
the definition of an hybrid approach of fraud detec-
tion that combines the characteristics of the canonical
non-proactive approaches with those of our proactive
approach.
ACKNOWLEDGEMENTS
This research is partially funded by Regione Sardegna
under project Next generation Open Mobile Apps
Development (NOMAD), Pacchetti Integrati di
Agevolazione (PIA) - Industria Artigianato e Servizi -
Annualit`a 2013.
REFERENCES
Attenberg, J. and Provost, F. J. (2010). Inactive learn-
ing?: difficulties employing active learning in prac-
tice. SIGKDD Explorations, 12(2):36–41.
Bhattacharyya, S., Jha, S., Tharakunnel, K. K., and West-
land, J. C. (2011). Data mining for credit card fraud:
A comparative study. Decision Support Systems,
50(3):602–613.
Bowyer, K. W., Chawla, N. V., Hall, L. O., and Kegelmeyer,
W. P. (2011). SMOTE: synthetic minority over-
sampling technique. CoRR, abs/1106.1813.
Breiman, L. (2001). Random forests. Machine Learning,
45(1):5–32.
Brown, I. and Mues, C. (2012). An experimental compari-
son of classification algorithms for imbalanced credit
scoring data sets. Expert Syst. Appl., 39(3):3446–
3453.
Chatterjee, A. and Segev, A. (1991). Data manipulation
in heterogeneous databases. ACM SIGMOD Record,
20(4):64–68.
Donmez, P., Carbonell, J. G., and Bennett, P. N. (2007).
Dual strategy active learning. In ECML, volume 4701
of Lecture Notes in Computer Science, pages 116–
127. Springer.
Duhamel, P. and Vetterli, M. (1990). Fast fourier trans-
forms: a tutorial review and a state of the art. Signal
processing, 19(4):259–299.
Faraggi, D. and Reiser, B. (2002). Estimation of the area un-
der the roc curve. Statistics in medicine, 21(20):3093–
3106.
Gao, J., Fan, W., Han, J., and Yu, P. S. (2007). A general
framework for mining concept-drifting data streams
with skewed distributions. In Proceedings of the Sev-
enth SIAM International Conference on Data Min-
ing, April 26-28, 2007, Minneapolis, Minnesota, USA,
pages 3–14. SIAM.
Japkowicz, N. and Stephen, S. (2002). The class imbal-
ance problem: A systematic study. Intell. Data Anal.,
6(5):429–449.
Lessmann, S., Baesens, B., Seow, H., and Thomas, L. C.
(2015). Benchmarking state-of-the-art classifica-
tion algorithms for credit scoring: An update of re-
search. European Journal of Operational Research,
247(1):124–136.
Phua, C., Lee, V. C. S., Smith-Miles, K., and Gayler, R. W.
(2010). A comprehensive survey of data mining-based
fraud detection research. CoRR, abs/1009.6119.
Pozzolo, A. D., Caelen, O., Borgne, Y. L., Waterschoot, S.,
and Bontempi, G. (2014). Learned lessons in credit
card fraud detection from a practitioner perspective.
Expert Syst. Appl., 41(10):4915–4928.
Pozzolo, A. D., Caelen, O., Johnson, R. A., and Bontempi,
G. (2015). Calibrating probability with undersampling
for unbalanced classification. In IEEE Symposium Se-
ries on Computational Intelligence, SSCI 2015, Cape
Town, South Africa, December 7-10, 2015, pages
159–166. IEEE.
Smith, S. W. et al. (1997). The scientist and engineer’s
guide to digital signal processing.
Sorournejad, S., Zojaji, Z., Atani, R. E., and Monadjemi,
A. H. (2016). A survey of credit card fraud detection
techniques: Data and technique oriented perspective.
CoRR, abs/1611.06439.
Vinciotti, V. and Hand, D. J. (2003). Scorecard construc-
tion with unbalanced class sizes. Journal of Iranian
Statistical Society, 2(2):189–205.
Wang, H., Fan, W., Yu, P. S., and Han, J. (2003). Mining
concept-drifting data streams using ensemble classi-
fiers. In Getoor, L., Senator, T. E., Domingos, P. M.,
and Faloutsos, C., editors, Proceedings of the Ninth
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, Washington, DC,
USA, August 24 - 27, 2003, pages 226–235. ACM.
