The document discusses Jordan Customs' need for a solution to help predict people and goods crossing borders that are most likely to break laws by smuggling or not paying proper customs duties, while not constraining legitimate trade. It describes building models using decision trees, naive Bayes, logistic regression, and association rules on customs declaration data to identify high-risk cross-border activities. Testing showed decision trees performed best. Recommendations include using longer study periods to determine persistent trends and building specific models for certain goods types or customs centers.
2. Jordan Customs & Boarder Control .
Formatting investment and enabling industry to be competitive to enhance the
competency of national economy.
Facilitating trade exchange between the Kingdom and the other countries.
Collecting revenues for the treasury.
Controlling passengers and goods movements and transportations crossing the
Kingdom’s borders in conformity with the departments authorities under the
current regulations in force.
Combating smuggling.
Protecting the local society and the environment from hazardons materials.
Contribution in controlling the commercial activities to prohibit illegal businesses
under the current regulations in force.
3. Problem Statement.
• The Customs department in Jordan operates in environments of uncertainty
and change. JC can’t manually check every person and goods that enter to the
country.
• In the past. 100 percent of the declarations examined, were everything that
entered the country was examined. Goods were processed in days, not hours,
and developing competent customs officers took years of training.
• The Asycuda system has been implemented in 2001 , it includes a risk
management modules that can randomly sample imports for compliance and
has some rules and lists of risks that are built based on the experience of the
JC officials.
4. Problem Statement.
• The Customs department in Jordan operates in environments of
uncertainty and change. JC can’t manually check every person and
goods that enter to the country.
• JC wants a way to predict the people and goods crossing a border who
will most likely breaking the law by smuggling drugs, or not paying the
right custom tariffs.
• At the same time the custom department shouldn’t constrain the
legitimate trade.
• JC needs to identify cross-border activities or transactions with the
highest potential to pose risk
5. The required Solution
Jordan Customs requested tp significantly improve
performance in facilitating trade and evidence
positive cost benefit outcomes through
automation, prediction , pre-arrival processing and
post clearance audit.
6. Mining Data Set
The dataset we used for this study covers the period
between January 2009 and June 2009, a list of the
constraints and assumptions we used to build the data
set is listed at the end of this presentation.
9. Input Variables
• Export Country( )
• Origin Country( )
• HSCode at level 11( )
• Company( )
• Clearance company( )
• Whether there is Exemption or not and in case there is exemption what kind of
exemption( )
• Whether there is Agreement or not and in case there is agreement what kind of
agreement( )
• Declaration Type(
• Custom Status( )
• Custom center ( )
• TotalitemTaxAmount ( )
We used the following variables to study the fraud bavior
10. Used Mining Algorithms
Decision Trees a classification data technique that splits data using
tree figure where it create nodes that contain the transactions that belongs
to certain state of an attribute based on the special characteristics that
differ a certain value of certain variable from the remaining dataset.
Naïve base a classification algorithm that looks at attributes and treat
them as if they are completely independent and looks at the effect of one
variable at time on the predicted output at its classification. It is powerful
in comparing the importance of each input variable and to know the most
importance ones.
The fooling algorithms where used
11. Logistic Regression :- a classification algorithm that represent a special
form of neural network. It is used in cases where output is one of two possible
states or goes on one of two directions only, and it is not recommended to be
used in cases where multiple classes are used. In case of multiple classes neural
networks can be used rather.
Association Rules :- an association algorithm that asses to what extent a
group of item sets (values or states that appear together in a certain number of
transactions) are associated to specific value of predicted variable
Used Mining Algorithms
17. Testing Results – One month later
We run the same
model for a
dataset of
declarations for a
one month later
and here are the
results
18. Conclusions and Results
Recommendations
The scores (importance results) of the LR were high for the attribute values that have very
few cases, and we recommend to not rely on the score values in creating risk rules.
Neural network can be used. However, it will share the same issues with LR (similar
behavior specifically for scaled networks). Besides, performance issues could be
encountered in using neural network (very long processing time and very long exploration
time).
In order to produce better results networks should be unscaled and could be achieved by
avoiding input attributes that have big number of possible sates such as HS code at level 11
or company name. It might be a good idea to try to use Chapters rather than HS codes at
level 11 in deeper studies if neural networks were chosen as a model.
A further study for neural network and logistic regression could be done by trying to create
a dataset that ignores input states that have few supporting cases for example less than 10.
19. Conclusions and Results Recommendations
- 2 -
Naïve base is good algorithm in studying one attribute at time and determining the most
important attributes but it should be known that it differs from other algorithms that it
could not study more than one attribute at time.
Decision tree is a very effective algorithm that a deep analysis of its details could represent
a good guider in identifying the patterns in data at customs and detect fraud behavior.
Association rules could be useful in identifying patterns in data however analysts should
take into consideration that it differs from classification techniques.
Generative algorithms such decision trees could be useful to classify datasets where the data
set related to one node could be used in doing further analysis using discriminative
algorithm such as neural network. Where it is important to emphasize that if the resulted
network is unscaled the result will be more useful.
If model cascading used and models need to be created from excel it is recommended to do
that in a separate analysis services database that could be used from excel. Where this
operation is recommended in order to avoid using built excel models in case reprocessing
took place from sql data tools.
20. Conclusions and Results Recommendations
– 3-
In this study we used data for six month. However, in further studies we recommend the
data set to be for longer periods. Where short periods has the advantage of detecting recent
trends longer periods is more accurate in determining persistent trends.
In this study we used the month just after the study (July 2009) where in further studies it is
a good idea to give a gap month for the analysis and to assess the model for a short period
after that. Also, it is important to emphasize that study periods should be much longer
than testing periods. For example, a two years study could be done and the results of the
two months following to the period of one month after the study could be taken and
statistics similar to the ones we did for July 2009 could be done to assess models reliability.
Also, it is recomended for the studies to be done by customs for different periods to assess
how stable is the results of the models in order to reach most suitable period for the study
and actions taken according to studies
In this study we took all the import declarations. Where in further studies, it is useful to try
to build a specific models such as models for certain chapters or models that eliminate
certain chapters. Also, studies could be done by customs for specific custom centers to get
insights about behaviors and patterns related to the specific center.
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset
. Association rules algorithm differs from the previous algorithms as it looks for associating data not classifying them as per the previous algorithms. In association the algorithm try to find whether there is a rule that contains certain state of attribute or set of states from different attributes that implies a certain state for the predictable attribute. It is mainly used in recommendation scenarios, it is more interpretable in viewing. However, we should not interpret the probability resulting from it in a way similar to the probability from previous algorithms where the probability here is more derived from the rule that might fire. So, it is not recommended to use it in quantifying risk associated to certain scenarios rather it could be used in better understanding the data and patterns related to the dataset