Detecting Shipping Fraud in Global Supply Chains using Probabilistic
Trajectory Classification
Ron Triepels
1
and Hennie Daniels
1,2
1
CentER, Tilburg University, Warandelaan 2, Tilburg, The Netherlands
2
Rotterdam School of Management, Erasmus University, Burg.Oudlaan 50, Rotterdam, The Netherlands
1 INTRODUCTION
Advances in transportation technologies and the liber-
alization of trade restrictions have transformed trade
into a global phenomenon during the last two decades
(Taylor, 2002). While the economic impact of global
trade has been widely studied, it is only recently that
researchers in the field of supply chain management
have begun to investigate the negative impact of glob-
alization on supply chain visibility. The inability to
track goods in transit involves risks, particularly in
terms of the customs brokerage practices of freight
forwarders.
Restrictions in international trade are fading away
due to preferred trade agreements between countries
and trade organizations. Firms within preferred trad-
ing areas can exploit differences in wages and knowl-
edge across countries. This allows them to acquire
difficult-to-obtain goods at competitive prices. How-
ever, the increasing diversity of firms, products, and
shipping options makes global supply chains consid-
erably complex. Trade is initiated between buyers and
sellers whose identity and trade behavior is relatively
unknown to the freight forwarder. Also, global supply
chains affect a lot of intermediate shipping companies
that operate in highly competitive markets. Shippers
are afraid to reveal too much information and be put
out of business by competitors.
At the same time, innovation and technology have
significantly reduced transportation costs (Hummels,
2007). One of the most noteworthy innovations in lo-
gistics is containerization. Containerization refers to
a system of containers designed to move goods ef-
ficiently across all modes of transport. Goods at the
beginning of the supply chain are clustered in contain-
ers based on their ownership or destination. The stan-
dardized design allows shippers to quickly exchange
containers at terminals without having to open them.
Although this greatly improves efficiency, it also neg-
atively affects supply chain visibility because what is
being shipped in a given container is no longer visible
from the outside.
Risks for the freight forwarder mainly arise when
goods marked for import must be declared to the cus-
toms authorities of the importing country. Several in-
termediate shippers may be involved in a single trans-
portation who only share a bill of lading to attest
to their provided services. The problem here is that
there is no guarantee that the goods listed on these
documents are actually inside the containers. Open-
ing every container marked for clearance is problem-
atic because this would imply major operational costs.
Instead, freight forwarders must solely rely on ex-
ternally created shipping documentation and declare
goods they usually do not even seen (Hesketh, 2010).
The rise of information technology provides op-
portunities to improve the customs brokerage prac-
tices of freight forwarders (Gordhan, 2007). On one
hand, information technology allows more detailed
supply chain data to be recorded. Radio frequency
identification (RFID) has received significant atten-
tion in the supply chain sector. The small size of
the RIFD tags and their low production costs makes
them useful in tracking and tracing international cargo
flows (Chang et al., 2011) and reducing delays at cus-
toms clearance locations (Hsu et al., 2009). On the
other hand, information technology provides the tools
to share data among supply chain participants. Sev-
eral technologies have been proposed to connect ship-
pers and freight forwarders, such as electronic data
interchange (EDI) (Murphy and Daley, 1999). This
project investigates how freight forwarders can use
data mining to extract knowledge from their extensive
supply chain repositories and use it to fight shipping
fraud.
2 STATE OF THE ART
Shipping fraud is committed in many different ways
and on different scales, ranging from local cargo theft
to international smuggling. Either way, evidence of
a fraud scheme must be covered in the corresponding
12
Triepels R. and Daniels H..
Detecting Shipping Fraud in Global Supply Chains using Probabilistic Trajectory Classification.
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
shipping documentation. This is known as document
fraud or misrepresentation. Misrepresentation is the
act of manipulating facts in contracts or agreements
with the intent to benefit by commercial gain (Hill and
Hill, 2009). The tendency of fraudsters to manipulate
documentation elicits an interesting and challenging
question. To what extent does shipping documenta-
tion correspond to reality? Researchers have studied
how different data mining techniques can help freight
forwarders answer this question.
2.1 Classification
Classification is the problem of finding a good model
that assigns an observation to the right class based on
a set of observations with known classes.
Classification can be used to determine whether
goods on a declaration are correctly classified (Filho
and Wainer, 2007). The researchers constructed a hi-
erarchical Bayesian model consisting of four features:
type of goods, consignee, country of origin, and desti-
nation country. Associations between the features are
learned from a large sample of correctly and incor-
rectly classified goods. It turns out that the proposed
model achieves high prediction accuracy despite the
high cardinality of the features.
A related study takes a broader approach and con-
structed a classifier based on association rules to pre-
dict the risk level of a declaration (Yaqin and Yum-
ing, 2010). In this study, the classification model is
built by performing association rule mining on a large
set of features. In addition to the model by Filho
and Wainer, this study also considers features such
as weights and prices. The risk level of a declaration
is determined by the number and nature of association
rules that match. Freight forwarders can use these risk
levels to pay more attention to declarations that have
a high probability of being fraudulent.
2.2 Regression
Regression is the problem of finding a good function
that captures the underlying relationship between a set
of variables.
Logistic regression has been applied to predict
the extent to which a declaration involves smuggling
(Hua et al., 2006). Fitting a good regression function
is difficult due to the inhomogeneous nature of cus-
toms data. The study uses a two-step cluster method
to divide customs data into seven approximately ho-
mogeneous clusters. Similarity between declarations
in these clusters is based on features such as the prices
and weights of goods. For each cluster, a logistic re-
gression function is fitted with a set of variables that
are correlated with smuggling. Finally, a decision rule
is defined to combine the predictions of the individual
regression functions. Performance tests show that the
decision rule significantly improves the efficiency of
customs inspections.
2.3 Anomaly Detection
Anomaly detection is the problem of identifying com-
binations of features that significantly deviate from a
statistical norm.
Brazil has built an anomaly detection system for
their customs systems in conjunction with universities
in the country (Digiampietri et al., 2008). The system
uses Markov chains and n-grams to search for product
descriptions on a repository of historic traded goods.
If a satisfactory match is found, historic data about
the product and its properties are retrieved and high-
lighted in a set of diagrams. These diagrams show
the statistical distribution of combinations of product
properties. In this way, customs agents can visually
inspect how much goods marked for clearance devi-
ate from what is expected.
Anomaly detection suffers from two major draw-
backs. First, it can be hard to pinpoint anomalies from
diagrams alone, as the decision boundary is defined
by a subjective interpretation. There is a general ten-
dency to overrate anomalies, which results in a over-
load of suspicious cases. Second, anomalies cannot
be compared based on their severity. This makes it
difficult to make a deliberate choice between the cost
to undertake a detailed investigation and the savings
that can potentially be achieved.
Ranking can be used to prioritize anti-fraud inves-
tigation (Kopustinskas and Arsenis, 2012). The study
proposes a way to calculate a numerical ranking for
price outliers in trade data. A suitable method is to
multiply the expected loss in duties by the probabil-
ity that a declaration is fraudulent. Expected loss is
estimated by multiplying the traded quantity and dif-
ference in unit price, whereas the fraud probability is
estimated by the p-value of statistical tests. The rank-
ing measure allows freight forwarders to identify dec-
larations for further investigation, which can achieve
high savings.
3 RESEARCH PROBLEM
Shipping fraud has mainly been investigated by an-
alyzing product specifications and aggregated geo-
graphical data. Research shows that particular types
of shipping fraud can be detected when looking at
the statistical properties of declarations. However,
DetectingShippingFraudinGlobalSupplyChainsusingProbabilisticTrajectoryClassification
13
these approaches mainly neglect how goods actually
find their way to the consignee. The trajectory of
a shipping container presumably reflects hidden pat-
terns that are characteristic of certain types of fraud.
The current literature lacks a well-grounded explana-
tion of how fraud influences shipping trajectories and
of how we can detect it from supply chain data.
International trade is moving towards vertical spe-
cialization in which each country produces particular
goods for the stages of a production sequence (Hum-
mels et al., 2001). At the same time, firms seek ways
to optimize the logistics between countries based on
economic considerations, such as price, flexibility,
and service level (Tongzon, 2009; Chang et al., 2011).
Vertical specialization and shipping optimization is
expected to create distinct patterns that reflect how
countries trade and under what conditions goods are
being transported. Deviations from what are expected
to be normal shipping trajectories can therefore point
to possible cases of fraud.
3.1 Trajectory Classification
Trajectory classification is the problem of predicting
the class of moving objects based on historic trajecto-
ries (Lee et al., 2008). This problem has recently re-
ceived increased attention in the data mining field due
to its potential applications. Trajectory classification
is applied in video surveillance to detect anomalous
behavior (Owens and Hunter, 2000) and air pollution
measurement to study the pollution of air masses ap-
proaching a given site (Riccio et al., 2007). Despite
the increased attention, the application of trajectory
classification in global supply chains has for the most
part been unexplored.
The classification of moving objects can be prob-
abilistic or non-probabilistic. A probabilistic classi-
fier uses a generative model to determine the most
likely class by calculating a probability distribution
over all classes. In contrast, a non-probabilisticclassi-
fier predicts a class without specifying the uncertainty
by which the prediction is made. The use of a prob-
abilistic classifier has the advantage that it allows the
incorporation of decision theory and the making of
automated decisions based on utility functions. De-
cision theory is vital for fraud detection systems to
prioritize anti-fraud investigations. This leads to the
following research question.
Research Problem. How can freight forwarders ap-
ply probabilistic trajectory classification to predict
shipping fraud in global supply chains?
We identified four sub-problems that need to be
studied to apply trajectory classification to supply
chain data. This includes the generation of discrimi-
nating patterns, the support of multidimensional data,
the reliability of model parameters, and the choice of
the probabilistic classifier.
3.2 Discriminating Patterns
An important process in the construction of an ef-
fective classifier is the generation of discriminating
patterns. Discriminating patterns are combinations of
features that occur with disproportionate frequency in
some classes (Fang et al., 2011). Including them in a
classifier has proven to significantly increase predic-
tive power. However the features of a shipping trajec-
tory are different from traditional features. First, lo-
cations of a trajectory occur in a logical order. Mov-
ing goods to a given location can be discriminatory,
but this premise may not hold for trajectories in the
reverse direction. Second, trajectories can become
quite large in size and can involve many different lo-
cations. This is expected to raise some performance
issues. Research is needed to effectively find discrim-
inating patterns in global supply chains.
Sub-problem 1. How can discriminating patterns in
shipping trajectories be effectively generated?
3.3 Multidimensional Data
Trajectory classification models in the literature are of
one dimension, that is, they only model the direction
in which an object is moving. While this is sufficient
in most applications, the movement of goods in a sup-
ply chain is influenced by many more dimensions or
shipping concepts. Take the International Commer-
cial terms or Incoterms as an example. Incoterms
were introduced to clearly communicate how tasks,
costs, and risks are allocated between the supply chain
participants. Because each supply chain participant
has a different stake in the shipment, different terms
likely imply different trajectories. An important ques-
tion is therefore how such additional dimensions can
be added to the classification of a trajectory.
Sub-problem 2. How can trajectories be classified
based on multidimensional trajectory data?
3.4 Reliable Parameters
The parameters of a probabilistic classifier need to be
estimated from data. This is a difficult task because
samples of real life data often include rare events
and are loosely controlled. Bayesian inference suf-
ficiently estimates these rare events by incorporating
prior knowledge and updating a distribution when ev-
idence is gleaned form a sample. Research argues that
ICEIS2015-DoctoralConsortium
14
Bayesian inference supports the most optimal way of
decision making (Braithwaite et al., 2007). However,
the problem that follows is how much, and from what
source, expert knowledge is needed to estimate the
model parameters in a reliable way. This problem
needs to be further explored to make classification ef-
fective in a practical environment.
Sub-problem 3. How can Bayesian inference be ap-
plied to estimate the parameters of the trajectory clas-
sifier in a reliable way?
3.5 Generalizability
The literature provides many techniques to perform
probabilistic classification, including naive Bayes and
tree-augmented Bayes (Friedman et al., 1997). The
choice of a classifier depends greatly on the nature of
the data from which the classification task is learned.
Naive Bayes suffers from the strong independence
assumption between features. This assumption can
partly be avoided by using a classifier such as tree-
augmented Bayes. However, performing model aver-
aging to find the best tree-like model remains difficult
(Cerquides and De M`antaras, 2003). It is important
to investigate the strengths and weaknesses of differ-
ent probabilistic classifiers and their generalizability
to typical supply chain data.
Sub-problem 4. Which type of probabilistic classi-
fier generalizes well to supply chain data?
4 OUTLINE OF OBJECTIVES
The primary goal of this research project is to build
a framework for freight forwarders to construct effec-
tive classifiers for the purpose of predicting fraudulent
shipping trajectories. Secondary goals are to:
1. Create a general description of the customs bro-
kerage processes of freight forwarders, including
tasks, actors, and information flows.
2. Identify the relationship between shipping trajec-
tories and different types of fraud.
3. Define a general process that describes how tra-
jectory classification can be set up, including:
(a) A strategy to effectively generate discriminat-
ing shipping trajectories.
(b) A method to support multidimensional trajec-
tory features.
(c) A strategy to estimate the model parameters in
a reliable way.
(d) An overview of different classifiers and their
expected performance.
4. Define a set of design principles that freight for-
warders can use to implement probabilistic trajec-
tory classification in practice.
5 METHODOLOGY
During this research project, a designable artifact will
be created that follows the design process proposed
by the design science paradigm. Design science is
defined as a research process that involves the cre-
ation and evaluation of information technology (IT)
artifacts designed to solve a specific organizational
problem (von Alan et al., 2004).
Our project is positioned between the environment
of freight forwarding and the knowledge base of the
data mining community. Freight forwarding practices
are combined with established techniques in data min-
ing to come up with a solution to the growing visibil-
ity problems in global supply chains. The designable
artifact will be a data mining framework that aims to
improve the decision making in customs brokerage by
analyzing historic shipment trajectories. To safeguard
validity and effectiveness, we design the framework
according to a modified version of the design science
research model (Peffers et al., 2006).
Research activities are divided into four design
stages; see Figure 1. In the first stage, we explore
the problem that occurs in the domain of interest
and define a corresponding objective for the research
project. Both topics have already been discussed in
previous sections. Therefore, the remainder of this
section elaborates on the other stages of the research
project.
5.1 Requirements
In the second stage, the requirements that need to be
satisfied by the data mining framework are defined
by performing a requirements analysis. The require-
ments analysis consists of a field study and a literature
review.
A field study will be performed at a large inter-
national freight forwarder to explore its customs bro-
kerage practices and corresponding information pro-
visioning. Customs brokerage has become a complex
process with the establishment of numerous interna-
tional trading policies and shipping concepts. There-
fore, this study is also conducted to make us familiar
with this domain knowledge. Emphasis will be put on
the way that shipping fraud is committed in practice
and howfraud characterizes itself in the data of a ship-
ment and its declaration. Data for the field study will
DetectingShippingFraudinGlobalSupplyChainsusingProbabilisticTrajectoryClassification
15
Stage 1
Problem
Lack of supply
chain visibility
Objective
Fraud detection
using trajectory
classification
Stage 2
Requirements
Field test
Literature
review
Stage 3 Stage 4
Design
Prototyping
Evaluation
Classifier
evaluation
Communication
Scholar
publications
Demonstration
Proof of
concept
Figure 1: The four stages of the research project (stage one is discussed in this paper and is therefore highlighted). The
template for this model is adopted from (Peffers et al., 2006).
mainly be collected though interviews with managers
and customs agents.
Existing solutions proposed in the data mining lit-
erature will be explored by an extensive literature re-
view. The literature review will be conducted to gain
an overview of existing solutions proposed to fight
shipping fraud. Papers dedicated to this topic are hard
to find and are scattered in various disciplines, in-
cluding data mining, customs, finance, and operations
research. Therefore, a comprehensive bibliography
search will be performed across various libraries and
journals. The literature search will only select con-
ference papers and journal papers. Each publication
will be classified according to the six main tasks of
data mining (Fayyad et al., 1996), along with the ad-
vantages and disadvantages of the chosen technique.
Solutions are compared, and the review will elaborate
on the extent to which these solutions are able to im-
prove supply chain visibility in practice. Insights will
be combined with the conclusions of the field study to
formulate the design requirements.
5.2 Design
In the third stage, the design of the data mining frame-
work takes places. Design activities include the con-
struction of classifiers using prototyping and the eval-
uation of these classifiers by conducting performance
evaluations.
Prototyping is the process of developing a set of
trial versions of an artifact to clarify requirements
and reveal critical design decisions (Gordon and Bie-
man, 1995). The technique has become increasingly
popular in the field of software development because
it allows the building of systems in a short span of
time with feedback from end-users. Prototypes will
be built by taking samples from the databases of the
freight forwarder. These samples will initially be rel-
atively small and concentrated in specific trade lanes
but will gradually increase in size and scope when
the project progresses. Each sample will be stored in
a Microsoft SQL server instance to explore the data
and perform data cleaning and transformation. Data
is then loaded into statistical package R to construct
different classifiers.
5.3 Evaluation
Each classifier that is built as a prototype will be
tested on its performance, that is, its ability to detect
shipping fraud. A classifier performs well if it scores
high on all of the following three quality indicators:
1. Accuracy - is the ability to correctly predict class
labels on previously unseen data. Accuracy will
be tested by separating data into a training set and
a test set and measuring the number of correctly
predicted classes in the test set.
2. Scalability - is the ability to deal with large data
sets within an acceptable amount of time. Scala-
bility will be tested by increasing the sample size
and scope while measuring the effect on the accu-
racy and time complexity.
3. Robustness - is the ability to make correct predic-
tions based on loosely controlled data. Robust-
ness will be tested by decreasing the amount of
prior information used during parameter estima-
tion and measuring the effect on the accuracy.
The performance of all classifiers in a design it-
eration are compared to each other. If there exists a
classifier that performs satisfactory, i.e. it scores at
least high on accuracy, then a proof of concept will be
ICEIS2015-DoctoralConsortium
16
built. Otherwise, the design requirements are revised
in close corporation with the freight forwarder.
5.4 Demonstration
In stage four the results of and insights about the de-
sign activities are reported to the freight forwarder
and the data mining community.
The best performing classifier will be used to built
a proof of concept. A proof of concept will be build to
demonstrate the main features of the classifier to the
freight forwarder and to verify that it has real-world
application. The demonstration will show how the
classifier can be used during the customs brokerage
processes, how employees can operate the classifier,
and what savings the freight forwarder can expect to
achieve. If the freight forwarder accepts the proof of
concept, then the data mining technique and the proof
of concept will be reported to the data mining com-
munity via scholarly publication. Otherwise, the de-
sign requirements will be revised and a new proof of
concept will be built.
5.5 Communication
Important findings in the design of an effective classi-
fier will be reported back to the data mining commu-
nity through scholar publication. The research project
is likely to produce at least the following papers:
1. A paper on the motivation and possibilities to use
supply chain trajectoriesto identify different types
of shipping fraud.
2. A paper that describes the data mining framework
to apply trajectory classification in practice.
3. A paper with a case study that describes a proof
of concept applied at the freight forwarder.
The second paper may contain too much detail for
a single publication. Therefore, the paper is likely to
be divided into several smaller papers that address one
or two sub-problems and one that provides a general
overview of the framework.
6 STAGE OF THE RESEARCH
The first two stages of the research project are al-
most completed. An international freight forwarder,
whose identify will remain unknown for privacy rea-
sons, has been contracted for the project. Several in-
terviews havebeen performed with customs managers
and customs agents at different offices of the freight
forwarder. Based on these interviews, the customs
brokerage processes are described in detail. Also, the
literature review is completed from which the most
important publications are included in this paper. In-
sights from the interviews and the literature review
are used to define the first draft of the design require-
ments.
Regarding the third stage, samples are taken from
the shipping system and customs system of the freight
forwarder. Neither information system is currently in-
tegrated, so some effort was needed to correctly link
shipments and declarations in a new database.
The data is used to construct a Bayesian network
that models the locations of a cargo trajectory together
with a set of important declaration features (Liu et al.,
2014). Predictive reasoning is applied to investigate
the change in probabilities of features when evidence
of a trajectory is inserted into the network. Based on
the evidence inserted, the network predicts the type of
goods, the use of preferential documents, and the cus-
toms duties to be paid. The initial prototype demon-
strates the ability to detect human errors and poten-
tial cases of fraud by investigating the trajectory of a
shipment. However, we found the trajectory of a ship-
ment cannot be modeled on a sufficient granularity in
a Bayesian network. It is expected that more detailed
locations of the trajectory need to be included in the
classification model to achieve sufficient performance
for real-life application.
Currently, we are investigatinghow cargo trajecto-
ries can be modeled in more detail using discriminat-
ing sequential patterns. We use the three-step mining
approach proposed by (Cheng et al., 2007) to build a
classification model. First, sequential mining is per-
formed to find frequent combinations of locations in
the cargo trajectories. Second, feature selection is
performed to select the most discriminating patterns.
Third, the data set is transformed and a classifier is
built based on the selected features. The main ad-
vantage of this approach is that trajectories can be of
any size and can contain repeating parts of sequences.
Different classifiers and strategies to find discriminat-
ing trajectories are currently being investigated.
7 EXPECTED OUTCOMES
The expected outcome of this research project is a
framework for freight forwarders to construct effec-
tive trajectory classifiers for the purpose of detecting
fraud in international cargo flows. Freight forwarding
practices are integrated with established techniques in
the field of data mining. The resulting framework is
likely to impact both freight forwarding practices and
techniques for trajectory classification.
DetectingShippingFraudinGlobalSupplyChainsusingProbabilisticTrajectoryClassification
17
On one hand, freight forwarders struggle with the
extensive amount of data they have to deal with dur-
ing their operations. The framework guides freight
forwarders through the process of setting up data min-
ing initiatives and shows how such initiatives can im-
prove customs brokerage practices. Implementations
of the framework are likely to increase the effective-
ness of customs compliance. On the other hand, the
literature on trajectory classification lacks techniques
to analyze global supply chains. Supply chain data
has the characteristics of big data and is therefore dif-
ficult to analyze. The framework is expected to con-
tribute a new application of trajectory classification to
the data mining literature and show how it can effec-
tively be used vis-`a-vis on supply chain data.
REFERENCES
Braithwaite, R. S., Roberts, M. S., and Justice, A. C.
(2007). Incorporating quality of evidence into deci-
sion analytic modeling. Annals of internal medicine,
146(2):133–141.
Cerquides, J. and De M`antaras, R. L. (2003). Tractable
bayesian learning of tree augmented na¨ıve bayes mod-
els. In ICML, pages 75–82.
Chang, Y. S., Son, M. G., and Oh, C. H. (2011). De-
sign and implementation of rfid based air-cargo mon-
itoring system. Advanced Engineering Informatics,
25(1):41–52.
Cheng, H., Yan, X., Han, J., and Hsu, C.-W. (2007). Dis-
criminative frequent pattern analysis for effective clas-
sification. In Data Engineering, 2007. ICDE 2007.
IEEE 23rd International Conference on, pages 716–
725. IEEE.
Digiampietri, L. A., Roman, N. T., Meira, L. A., Ferreira,
C. D., Kondo, A. A., Constantino, E. R., Rezende,
R. C., Brandao, B. C., Ribeiro, H. S., Carolino, P. K.,
et al. (2008). Uses of artificial intelligence in the
brazilian customs fraud detection system. In Proceed-
ings of the 2008 international conference on digital
government research, pages 181–187. Digital Govern-
ment Society of North America.
Fang, G., Wang, W., Oatley, B., Van Ness, B., Steinbach,
M., and Kumar, V. (2011). Characterizing discrimina-
tive patterns. arXiv preprint arXiv:1102.4104.
Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P.
(1996). From data mining to knowledge discovery in
databases. AI magazine, 17(3):37.
Filho, J. and Wainer, J. (2007). Using a hierarchical
bayesian model to handle high cardinality attributes
with relevant interactions in a classification problem.
In IJCAI, pages 2504–2509.
Friedman, N., Geiger, D., and Goldszmidt, M. (1997).
Bayesian network classifiers. Machine learning, 29(2-
3):131–163.
Gordhan, P. (2007). Customs in the 21st century. World
Customs Journal, 1(1):49–54.
Gordon, V. S. and Bieman, J. M. (1995). Rapid prototyping:
lessons learned. IEEE software, 12(1):85–95.
Hesketh, D. (2010). Weaknesses in the supply chain: who
packed the box. World Customs Journal, 4(2):3–20.
Hill and Hill (2009). Nolo’s Plain-English Law Dictionary.
Nolo.
Hsu, C.-I., Shih, H.-H., and Wang, W.-C. (2009). Apply-
ing rfid to reduce delay in import cargo customs clear-
ance process. Computers & Industrial Engineering,
57(2):506–519.
Hua, Z., Li, S., and Tao, Z. (2006). A rule-based
risk decision-making approach and its application in
china’s customs inspection decision. Journal of the
Operational Research Society, 57(11):1313–1322.
Hummels, D. (2007). Transportation costs and international
trade in the second era of globalization. The Journal
of Economic Perspectives, pages 131–154.
Hummels, D., Ishii, J., and Yi, K.-M. (2001). The nature
and growth of vertical specialization in world trade.
Journal of international Economics, 54(1):75–96.
Kopustinskas, V. and Arsenis, S. (2012). Risk analysis ap-
proaches to rank outliers in trade data. In Advanced
Statistical Methods for the Analysis of Large Data-
Sets, pages 137–144. Springer.
Lee, J.-G., Han, J., Li, X., and Gonzalez, H. (2008).
Traclass: trajectory classification using hierarchical
region-based and trajectory-based clustering. Pro-
ceedings of the VLDB Endowment, 1(1):1081–1094.
Liu, L., Daniels, H., and Triepels, R. (2014). Auditing data
reliability in international logistics: An application of
bayesian networks. In Proceedings of the 16th Inter-
national Conference on Enterprise Information Sys-
tems (ICEIS 2014).
Murphy, P. R. and Daley, J. M. (1999). Edi benefits and
barriers: comparing international freight forwarders
and their customers. International Journal of Physi-
cal Distribution & Logistics Management, 29(3):207–
217.
Owens, J. and Hunter, A. (2000). Application of the self-
organising map to trajectory classification. In Visual
Surveillance, 2000. Proceedings. Third IEEE Interna-
tional Workshop on, pages 77–83. IEEE.
Peffers, K., Tuunanen, T., Gengler, C. E., Rossi, M., Hui,
W., Virtanen, V., and Bragge, J. (2006). The design
science research process: a model for producing and
presenting information systems research. In Proceed-
ings of the first international conference on design sci-
ence research in information systems and technology
(DESRIST 2006), pages 83–106.
Riccio, A., Giunta, G., and Chianese, E. (2007). The appli-
cation of a trajectory classification procedure to inter-
pret air pollution measurements in the urban area of
naples (southern italy). Science of the Total Environ-
ment, 376(1):198–214.
Taylor, A. M. (2002). Globalization, trade, and develop-
ment: Some lessons from history. Technical report,
National Bureau of Economic Research.
Tongzon, J. L. (2009). Port choice and freight forwarders.
Transportation Research Part E: Logistics and Trans-
portation Review, 45(1):186–195.
ICEIS2015-DoctoralConsortium
18
von Alan, R. H., March, S. T., Park, J., and Ram, S. (2004).
Design science in information systems research. MIS
quarterly, 28(1):75–105.
Yaqin, W. and Yuming, S. (2010). Classification model
based on association rules in customs risk manage-
ment application. In Intelligent System Design and
Engineering Application (ISDEA), 2010 International
Conference on, volume 1, pages 436–439. IEEE.
DetectingShippingFraudinGlobalSupplyChainsusingProbabilisticTrajectoryClassification
19
