See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/341741423

Data Mining for Mobile Internet Traffic Flow Forecasting

Conference Paper · May 2020

CITATIONS READS
0 95

3 authors, including:

Saber Khaled Elmabrouk Alhasan Salem

Libyan Academy Libyana mobile phone
22 PUBLICATIONS   34 CITATIONS    1 PUBLICATION   0 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

An Overview of Power Plant CCS and CO2-EOR Projects View project

An Overview of Power Plant CCS and CO2 -EOR Project View project

All content following this page was uploaded by Saber Khaled Elmabrouk on 29 May 2020.

The user has requested enhancement of the downloaded file.

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Data Mining for Mobile Internet Traffic Flow Forecasting

Saber Kh. Elmabrouk
School of Applied Sciences and Engineering, Libyan Academy
Janzur, Tripoli, Libya
[email protected]

Alhasan Nureddin Salem

Libyana Mobile Phone, Tripoli, Libya

Abstract
Internet traffic can be described as a general term that includes the transmission of internet data between different
devices and systems. The analysis and prediction of internet traffic is a proactive approach to ensure secure, reliable
and qualitative network communication. Several linear and non-linear models are proposed and tested to analyze the
prediction of network traffic, including techniques based on regression analysis, artificial intelligence and data mining.
These interesting combinations of internet traffic analysis and forecasting techniques are implemented to achieve
efficient and effective results. Timely and accurate prediction of the use of internet data is a topic of great importance
in the telecommunications industry. In addition, internet traffic data is important for many applications in
telecommunications management, such as understand customer behavior, optimal planning of the capacity of
networks, successful decision making and maintaining the quality of services at guarantee level in the future. This
situation inspires us to rethink data mining with internet traffic forecasting problems. This study provides an overview
of data mining in telecommunications and proposes a novel model for forecasting mobile internet traffic based on
artificial neural networks. The analysis of mobile Internet traffic data during the last five years shows that the month
of August has a higher traffic flow, while the lowest flow was the month of June. The selected model has three layers
with a determination coefficient of 97.5%.

Keywords
Data mining, Internet traffic, Forecasting, Telecommunication, ANN.

1. Introduction
Each manager wants to know the exact nature of future events, act accordingly, or plan activities when there is
sufficient time to execute the plan. The effectiveness of the plan depends on the awareness of future events. But all
managers are planning for the future, whether or not future events are known. This means that they are trying to predict
the future based the historical data and on their judgment and experience. In telecommunications companies, the
managers using historical data to predict the future to understand customer behavior, maintain quality of service,
optimize network capacity, monitor network traffic flow, network failures, maximize profit, fraud detection, anticipate
new products and services to improve existing products and services, and much more. Accurate estimation of network
performance is critical to the success of any type of network, whether voice or data. The two most important factors
that it needs to consider in managing the telecommunications are service and cost. Quality of service is essential to
maintain customer satisfaction whilst cost always influences profitability. This study aims to provide an overview of
data mining in telecommunications and proposes a novel model for forecasting mobile internet traffic based on
artificial neural networks (ANNs)

2. An Overview
Recently, the concept of data mining has been increasingly used in various industries. This discipline is primarily
concerned with designing algorithms to extract new information from large, existing structured data. Data mining,
also known as Knowledge Discovery in Databases (KDD). KDD was coined in 1989 by Gregory Piatetsky-Shapiro
during the first workshop on KDD (KDD-89). The term data mining appeared in the database community in 1990
(Fayyad et al., 1996). However, the challenge of data mining is to draw valuable analytical conclusions. Thus, most
telecommunications companies have realized that the vast volume of data they collect and possess could be effectively
utilized for solving their business problems by converting them into information and knowledge. Data mining can be
considered as a technique that automatically generates this knowledge from available data (Joseph, 2013). As a result,

© IEOM Society International 3071

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

many authors have analyzed the use of telecommunication services in different contexts, but little research has focused
on the data traffic generated, particularly on smartphones and their ability to use them.
The methods used in data mining, such as machine learning (ML), belong to the field of artificial intelligence (AI). A
system becomes artificially intelligent by feeding it with relevant patterns. ML is a set of algorithms that analyze data
and learn from them to make informed decisions. Whereas ANNs is one such group of algorithms for ML. Thus, ANN
is a special set of algorithms that has revolutionized ML. However, as ANN best identify patterns or trends in data,
they are well suited for prediction and forecasting applications. Anyway, different AI techniques have been
successfully applied to a wide range of extraordinary activities in the telecommunications industry. Previous research
has shown that data mining has used various types of telecommunication data that can be classified into three different
sets, namely: call details, network and customer data.

Call details data: This is information related to calls, which is stored as a record of the details of the call that contains
information such as originating and terminating phone numbers, date, time and duration of call. Usually this dataset
is not directly used for data mining (Joseph & Madhuri, 2013). The following features can be generated from this data;
Average call duration, Average number of call created per day, Average number of call received per day, Percentage
of no-answer calls, Percentage of day time calls, and Percentage of weekday call.

Network data: Telecommunication networks contain thousands of interconnected components. These components are
capable of generating big data. This data is used for network management such as network diagnosis and fault
detection. Data mining technologies are used to identify network faults by automatically extracting knowledge from
this dataset. Chaudhary & Singh (2015) defined the network fault as an abnormal operation or defect at the component,
equipment, or sub-system level that is significantly degrades performance of an active entity in the network or disrupts
communication.

Customer data: Since telecommunications companies have millions of customers, it is very important to have a
database to store information about their customers. This information may include; customer name, address, billing
data and payment history, call detail data, subscription information, service plan, and more.

To survive in a highly competitive telecommunications market, companies need to adapt to external changes.
Customer Relationship Management (CRM) has proven to be an important control tool for identifying environmental
changes in business processes and making difficult decisions. The literature review shows that most data mining
applications in telecommunications fall into three categories: CRM, fraud detection, and network management.
However, most of these studies were devoted to customer relationship management. These three categories are
described below with a general description of their application in data mining.

Customer Relationship Management (CRM): A CRM is defined as a set of activities a business performs to identify,
qualify, acquire, develop and retain loyal and profitable customers by delivering the right product or service to the
right customer, through the right channel, at the right time, and the right cost (Galbreath & Rogers, 1999). Data mining
applications in CRM can take many forms, among them; 1) Basket Analysis: developing inventory and store layout
strategies. 2) Sales Forecasting: examines time-based patterns for make restocking decisions, helps in supply chain
management, and financial management. 3) Database Marketing: increasing productivity, optimization company
resources based on customer profile, demographics, tastes, preferences, and purchasing behavior, and more. 4)
Predictive Life-Cycle Management: helps the organization predict customer lifetime and properly serve each segment.
5) Market Segmentation: increase efficiency by finding out which customers are interested in buying your products
and designing marketing campaigns and promotions based on their tastes and preferences. 6) Product Customization:
predict which features need be bundled to meet customer demand by customize products according to the exact needs
of customers, and 7) Fraud Detection: By analyzing what has turned out to be fraudulent in the past, the company can
take corrective action and stop the future occurrence of such events. Thus, the lessons learned and collected data from
customer information should be used to improve business and customer relationship behavior.

Moreover, nowadays, the telecommunication market matured and the churn rates became high. This maturing of the
market and increasing competition leaded the companies to focus on their existing customers and to find a way not to
let them go. Note that the term "churn" refers to the monthly or annual turnover of the clientele as defined by Strouse
(1999). Customer churn rate designates the percentage of customers the company has lost over a given time period.
Churn rate is a reverse side of retention rate, which shows the percentage of customers a company has retained over a
specific period. The most practical ways to calculate customer churn rate is; Customer Churn Rate = (Customers lost

© IEOM Society International 3072

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

during the period/Total customers at the start of the period) *100, where: Customers lost = Total Customers at the
start of the period + New customers during that period - Total Customers at the end of a period (Bitrix24, 2019).
In recent years, researchers have focused on CRM applications such as; customer value and retain profitable customer,
generate customer profitable and mine these for marketing purposes, acquire new customer, maximize profit, and
churn analysis which includes churn prediction and churn management, among of them; Masand et al (1999), Mani et
al (1999), Mozer, et al (2000), Adomavicius & Tuzhilin (2001), Cox (2002), Wei & Chiu (2002), Rosset et al (2003),
Hwang et al. (2004), Al-Rababah et al (2004), Freeman & Melli (2006), Hung et al. (2006), Camilovic (2008),
Haridasan & Venkatesh (2011), Khaligh et al (2012), Agarwal & Ranjan, (2013), Ye et al (2013), Arumawadu et al
(2015), Tripathi et al (2018), Halibas et al (2019), and Velayati et al (2019).

Fraud Detection: More than a third of organizations have experienced economic crime as reported by over 6,000
respondents to PwC’s Global Economic Crime Survey in 2016. In many existing literature, subscriber intent plays a
central role in defining fraud. Johnson (1996) defines fraud as any transmission of voice or data across a
telecommunications network where the intent of the sender is to avoid or reduce legitimate call charges. Also, Kaski
et al (1998) defines fraud as obtaining unbillable services and undeserved fees. Likewise, Blavette (2001) noted that
telecommunication fraud occurs whenever a perpetrator uses deception to receive telephony services free of charge or
at a reduced rate. It is a worldwide problem with substantial annual revenue losses for many companies.

There are many fraud techniques, some of which are quite sophisticated and combine more than one known method.
Telecommunications fraud is not static; new techniques evolve as the companies put up defenses against existing ones.
The fraudsters are smart opponents, continually looking for exploitable weaknesses in the telecom infrastructure. Part
of their motivation is accounted for by the fact that once an exploit is defined, there are thousands (or millions) of
potential targets. New types of fraud appear regularly, and these schemes evolve and adapt to attempts to stop them
(Becker et al, 2010). However, most of the data mining work done in detecting telecommunications fraud aims to
detecting or preventing fraud (Fawcett & Provost, 1997 and Hilas, 2009) and subscription (Estevez et al, 2006 and
Farvaresh & Sepehri, 2011) methods, because these are the fraud methods that have caused the most damage to the
telecommunications industry. Some papers, such as Collins (1999 A), Collins (1999 B), Collins (1999 C) and Hoath
(1998) describe state of telecommunication fraud. However, these papers do not discuss the details of the detection
process or the organized fraud model. In 2002, Bolton & Hand produced a report on a review of statistics and ML
tools as an effective technology for telecommunications fraud. Likewise, Becker et al (2010) discussed some of the
major fraud systems and their techniques. The study led to general conclusions about fraud detection. They specifically
support the use of simple, understandable models, effective use of visualization and a flexible environment, and
emphasize the importance of data management. In 2018, Zhou & Lin summarized the anti-fraud measures available
in the telecommunications industry. Besides, many data mining techniques exist in the literature for (a) supervised
learning based on training data of known fraud and legitimate cases and (b) unsupervised learning with data that are
not labeled to be fraud or legitimate. The literature shows that there are three data mining techniques used for fraud
analysis: Bayesian network, decision tree, and back-propagation. The Bayesian network is used for the classification
task. Classification determines which class a given data belongs to by defining predefined categorical classes. Decision
trees are used to create descriptive models. Descriptive models are created to describe the characteristics of the error.
Akhter & Ahamad (2012) discussed in detail the role of ANN in the prevention of telecommunications fraud. They
came to the conclusion that ANN is a better way to detect phone fraud because it incorporates adaptability, speed and
efficiency. Farvaresh & Sepehri (2011) used a hybrid approach consisting of a pre-processing, grouping and
classification phase. In the clustering phase, SOM and K-means were combined, and in the classification phase,
decision tree, ANN, and supporting vector machines were examined as a single classifier and bagging, boosting,
stacking, majority and consensus voting as ensembles were examined. The performance of all single and ensemble
classifiers was evaluated using various indicators and compared with statistical tests. It was shown that support vector
machines among single classifiers and boosted trees among all classifiers have the best performance in terms of various
metrics. Recently, Arafat et al (2019) suggested the use of various ensemble classifiers to overcome the highly biased
dataset and to facilitate more accurate classification. They found that the Extreme Gradient Boosting algorithm gives
the best results in terms of correctness and performance.

Network Management: Network monitoring is so essential because it has live insights. Having up to the minute
knowledge about what’s happening in the network is invaluable for any networking team. we are able to act when
problems occur without relying on end-user reports. The real added value of the AI network management tool is that
it intelligently automates tasks that are only becoming more general and complex. As network tools become more
sophisticated and accurate, it becomes increasingly difficult to effectively manage and manage a limited workforce.

© IEOM Society International 3073

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Network monitoring tools have begun to introduce AI and ML into network systems. In addition to recognizing and
alerting users to problems. However, AI can apply solutions to common problems without human intervention. AI
and ML allow network performance monitors to learn about common network issues. AI and ML don't just work on
its own. It needs to be trained to respond to events that happen on your network. This means that you can customize
AI to analyze and respond to event types. If you identify the same problem over and over again, AI can figure out the
best solutions to deal with them. Finally, it will provide enough data to make a decision. In addition to being able to
find network problems, AI is an incredible resource for preventing malicious problems before they damage the
network. AI can improve problems when they are noticed, allowing users to focus more on issues that require deeper
analysis and insight (Hein, 2019 and Enterprise Integration, 2019).

In 1993, Sasisekharan introduced a method to store data and identify trends in faulty networks. Using this approach,
a number of patterns have been identified on AT&T's worldwide network, which can be used by AT&T to improve
network reliability. In another work, Sasisekharan et al. (1996) provided an approach to maintain a massively
interconnected communications networks over time. This approach has been used to detect and predict chronic
transmission faults in AT&T's digital communications network. Liu (2014) presented a monitoring and analysis
system for large-scale networks based on Hadoop. Their results demonstrated that the system can efficiently processes
4.2 Tb of traffic data from 123 Gb/s links with high performance and low cost every day. In 2016, Qu et al introduced
a multilevel model mining architecture that supports automatic network management by discovering interesting
patterns from telecommunication network monitoring data. The proposed architecture leverages and combines
existing frequent item set discovery over data streams, association rule deduction, frequent sequential pattern mining,
and frequent temporal pattern mining techniques while also making use of distributed processing platforms to achieve
high-volume throughput.

3. Data Mining for Forecasting

Telecommunication service provides some kind of time series data, that is, data collected over time, including a time
stamp. In the context of forecasting, the decision maker has to find ways to get the value of big telecommunication
data. However, predictive data mining offers an opportunity to leverage the time series data sources that are already
available to decision makers that can directly impact profitability. For reference purposes, Rey et al. (2012) defined
short-range forecasts as one to three years, medium-range forecasts at three to five years and long-term forecasts at
more than five years. Generally, the authors agree that a period of more than ten years should be considered as a
scenario rather than a forecast. Most business modeling involves quarterly forecasting. Quarterly data is the frequency
with which the vast majority of external data providers store and forecast historical data. Gor (2003) defined
forecasting as the process of estimating a future event by transmitting historical data. historical data is systematically
combined in a predetermined manner to obtain an estimate of the future. Tuovila (2019) reported that forecasting is a
technique that uses historical data as inputs to make informed estimates that are predictive in determining the direction
of future trends. Businesses use forecasting to determine how their budget or plan their costs over the next period.
This is usually based on forecast demand for the products and services offered. Sanders (1998) has shown that poor
forecasting can lead to catastrophic decisions and prediction accuracy can make successful decisions, so forecasting
has improved in many areas, such as long-term forecasts by airlines. Whereas, prediction is concerned with estimating
the outcomes for invisible data. For this purpose, it fits a model into a training dataset, which results in an estimator
f^(x) that can make predictions for new x samples. Forecasting is a sub-discipline of prediction, in which we are
making predictions about the future based on time series. Thus, the only difference between prediction and forecasting
is that we consider the temporal dimension. The form of the forecasting is f^(x1, …, xt) where x1, …, xt indicate historic
measurements at time points 1,…,t, while the estimate relates to time point t+1 or some other time t in the future (Data
Science, 2019). Nonetheless, there are many ways to forecast. Each method is different and depends on the purpose
of the forecast, the data required, the availability of the data and the time frame for forecasting the event. Gor (2003)
presented three different forecasting methods; qualitative, time series, and causal methods.

3.1 The Process of Data Mining for Forecasting

Figure 1 summarized the seven steps of the data mining forecasting process; 1) identify the problem, 2) select the
forecasting period (time horizon), 3) select the forecasting method. This requires knowledge of the different
forecasting methods, which situations to use, how reliable they are; what type of data is needed. Based on these
considerations; one or more methods can be selected. 4) data selection: use the various metrics identified in step 2 as
an appropriate source that is compatible with the method(s) selected in Step 3. 5) make a forecast: apply the model
using the collected data and calculate the value of the forecast. 6) evaluation: before using the forecast model obtained

© IEOM Society International 3074

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

by one of the methods (step 5), it should be evaluated based of the confidence interval and error analysis. 7) the model
is ready to run.

Fig. 1 – The Seven Steps of the Data Mining Forecasting Process

4. ANN Applications for Telecommunications Forecasting

Recently, the use of ANN has contributed to improving the performance of the telecommunication service and defining
future technical and capacity requirements. This section provides a brief introduction to ANNs, followed by some
previous ANN studies on telecommunications applications. Basically, ANNs are used to model non-linear statistical
data and mimic the functioning of human brain neural networks. Figure 2 shows a simple neuron model.

Fig. 2 - A common Model for Neuron

It can be seen from Figure 2 that the neuron is a processing element that takes a number of inputs, x1, x2 ... xn, weighs
them, w1, w2, …., wn, and sum them up together with a bias parameter to get the value a. Then, a value is processed
by the activation function f(u) to get d, and then, the neural output error, e is calculated by subtracting the result d
from the target, t. The essence of the neuron model is to minimize the output error, according to some optimization
criteria, to improve the fit. Thus, all the neurons in the ANNs are trained using a block of input and output datasets so
that the NNs know the problem. The training can be classified into two groups: supervised and unsupervised training.
The former needs predefined training data that reflects network behavior. The network target is already known and
compared to the output when the input is applied to the network. Learning rule modifies w1, w2, …., wn and bias to
bring outputs closer to target. However, the weight (w1, w2, …., wn) and bias parameters are only updated with the
input data when using an unsupervised training. The multilayer perceptron neural network (MLPNN) is an extension
of the neuron model, and connections can only be one-way in the network. However, In the field of ANN modeling,
a radial basis function network (RBFNN) is an ANN activated by radial basis functions. It is often used to create
models for regression type problems (Pijush et al, 2017). Also, Group Method of Data Handling Polynomial Neural
Network (GMDH-PNN) is a model for obtaining high order input output relationship in time-series problems. GMDH-
PNN is an inductive unidirectional polynomial network that is made of large number of layers and each layer contains
many neurons (Koo et al, 2014).

Application of ANNs to traffic prediction by Ardhan et al (2007) Markus et al (2012), Raheem & Okereke (2014),
Ozovehe (2015) and Aliyu et al (2017) shows that ANN, RBFNN and GMDH-PNN can predict traffic with very high
accuracy. In 2004, Zhao et al. proposed two ANN models for network traffic forecasting to achieve two goals
forecasting a 24-hour load shape at a time and forecasting peak load per day. They collected network traffic data for
a year and a half. Data traffic from February 1, 1998 to July 31, 1999 was used as a training data set, and data from
August 1, 1999 to August 31, 1999 was used as a test data set. The 24-hour load shape prediction model consists of
four layers [24-20-30-1]. The input layer includes 24 inputs, two hidden layers with 20 nodes in the first hidden layer,
30 nodes in the second hidden layer, and one node in the output layer. The activation function used in the hidden layer
was a sigmoid, and the activation function for the output layer was a linear activation function. The peak load

© IEOM Society International 3075

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

forecasting model consists of five layers [16-15-20-30-1]. The input layer has 16 nodes, the output layer has only one
node connected with three hidden layers. The 1st hidden layer has 15 nodes, the 2nd hidden layer has 20 nodes, and the
3rd hidden layer has 30 nodes and trained by Levenberg-Marquardt algorithm. However, the names of the input
variables of both models were not mentioned in their study.

Cortez et al (2006) analyzed the effectiveness of three different time series forecasting approaches when applied to
TCP/IP internet traffic; Holt-Winters, ARIMA methodology and a Neural Network Ensemble (NNE) approach.
Comparison of these time series forecasting methods shows that NNE produced lowest errors. Data was collected
from two different ISPs in every five minutes and every hour. However, the input variables used in their study was
not specified. They used a logistic activation function in the hidden layer and a linear activation function in the output
layer. The Levenberg-Marquardt algorithm was used to construct the model. The average squared error was used to
measure the prediction performance of the models.

Oravec et al. (2008) proposed three different moving average neural network models for predicting video traffic
stream; multilayer perceptron, radial-based function networks and back propagation through time neural networks.
The multilayer perceptron network configuration was [3-10-1], the radial-based function network configuration was
[3-500-1], and the back propagation through time configuration was [3-20-1]. The activation function was logically
in the hidden layer and the linear activation function in the output layer. The Quasi-Newton algorithm was used to
train the model. The data used for training and the testing was taken from the video stream files of the
telecommunications network group. Training and testing dataset consists of 2000 points. The authors did not specify
the three input variables used in the first layer for the three proposed models. The best results of the forecast were
achieved from the back propagation through time network model.

In 2009, Gowrishankar & Satyanarayana modeled wireless network traffic in a short period of time using neural
networks and statistical methods. The results of both methods were compared at different time scales; 1 second, 10
seconds and 1 minute. The ANN model has a hidden layer, with a sigmodal activation function in the input layer and
a linear activation function in the output layer. They selected Quasi-Newton algorithm to train the model. The authors
did not specify the variables used in the first layer of the model.

Junsong et al (2009) proposed Elman neural network model to predict Internet traffic as a time series. The model
architecture was [2-2-1]. This means two variables in the first layer, one hidden layer with two nodes, and one target
(web traffic) in the output layer. Both input and output layers activated by sigmoid activation function. The training
algorithm was Levenberg Marquardt. The authors did not specify the two variables used in the model.

In 2010, Zaleski & Kacprzak introduced a Radial based neural network with a self-organizing algorithm that predicts
short-term values of traffic generated in packet switching networks. The selected model architecture [7-3-1] consists
of one hidden layer with three nodes activated with Gaussian activation function in the input layer. Whereas on the
output layer, one neuron acts as a linear activation function of hidden layer weighted outputs. The authors did not
define the seven variables used in the model as input layer.

Barabas et al. (2011) used two types of ANN (single-task learning (STL), multi-tasking (MTL) network approaches)
to predict the traffic load on Ethernet. The ANN model configuration consists of [4-5-1] for the STL approach and [8-
3-1] for the MTL approach. The activation function of neurons in the hidden layer was a sigmoid function, while the
output nodes have a linear transfer function. The Levenberg-Marquardt algorithm was the training algorithm. They
used 200 Mb consecutive traffic loads to build the prediction models, and 20 Mb values to test the models. They
concluded that the STL-based ANN model had the worst forecast performance, while the ANN model with the MTL
approach had the best performance. The authors did not define the selected variables used in the models as input layer.

In 2012, Botoca & Patrascu modeled two different approaches to ANN (Multilayer Perceptron (MLP) and Jordan-
Elman Neural Network) to predict channel quality in long-term evolutionary networks. Data was collected from radio
networks, 200 recorders were used to train the model, 77 records to validate it and 102 records to test it. The MLP
network structure consisted of [5-8-1]. Hyperbolic touch activation function used for all layers. The generated
propagation error algorithm used for training and validation. In the case of Jordan, the Elman NN structure was [4-8-
1], with a hyperbolic touch activation function in each layer. The generated propagation error algorithm used for
training and validation. They found that the Jordan-Elman Neural Network approach achieves better forecasting
results than the MLP network approach. The authors did not define the selected variables used in both models.

© IEOM Society International 3076

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Pandey et al. (2013) used ANN method to predict cellular network traffic during busy hours (between 5 pm and 7
pm). The ANN model consisted of [9-10-4-1]. Nine parameters were selected as input variables in ANN; SDCCH
Reservation Attempts, SDCCH Successful Calls, TCH Assignment, TCH Successful Calls, TCH Availability Ratio,
Inbound Transfer Success Rate, Outbound Transfer Success Rate, Mid-Term Traffic, and Average Hold Time. The
Tan sigmoid transfer function provided the best generalization of data across all layers using the Levenberg-Marquardt
as a training algorithm.

Ozovehe et al. (2018) formed three types of ANN (MLPNN, RBFNN, GMDH-PNN) and Neuro-fuzzy Inference
Systems models based on traffic measurements during busy hours taken from certain GSM / GPRS sites in Abuja to
predict traffic congestion in certain macrocells. They found that GMDH-PNN more predictive of TCH congestion
than other models as it resulted in lower errors for mean absolute error, standard deviation, root mean square error and
higher determination coefficient (R2). The authors did not define the selected variables used in the models.

Recently, Erunkulu et al. (2019) introduced an ANN model for predicting call drop events in a global network of
mobile networks. The accuracy of the model was 87.5%. The architecture of the model consisted of five layers [5-4-
5-4-1]. Inputs for the first layer are: received signal level, received quality, Image error rate, Bit error rate and Timed
transmission. However, the authors did not provide the training algorithm and activation functions used to build the
model.

5. Forecasting Mobile Data Traffic Over Mobile Phone Networks in Libya

5.1 Data Acquisition and Analysis
Mobile voice and internet data was recorded as a daily traffic monitor from January 1, 2013 to December 31, 2017 on
the Libyan real mobile network, which covers the Tripoli region. Between 2013 and 2017, it was noticed that the
demand for voice services increased insignificantly, while the demand for Internet services increased significantly.
Whereas the proliferation of smartphones supporting new mobile technologies has enabled mobile users to initiate
voice calls over Internet services, resulting in high data traffic of up to 1.5 Mb per user in 2017, thus reducing the need
for lower voice calls over the last five years from 2013 to 2017. Figure 3 shows a consumer comparison of voice and
data usage between 2013 and 2017.

Fig. 3 - Consumer Comparison of Voice and Data Traffic (2013-2017)

Nowadays, many applications such as voice, video calling, and social media have been replaced by voice over internet
services. The spread of modern mobile devices as the smartphone can support new mobile technologies. Today, the
behavior of mobile users has changed, and most of them are using smart devices to manage new mobile networks,
each of which increases data traffic and affects the network's ability to support traffic in the future. while demand for
voice services is expected to decline in the coming years. Table 1 shows that data traffic gradually increased by 93.2%
between 2013 and 2017. This indicates that the demand for Internet services will increase in the near future. This is a
challenge for telecom service providers to make decisions that can be used to avoid the risks of network congestion
and maintain quality of service. Figure 4 shows the distribution of mobile traffic data over the last five years (2013-
2017)
Table 1 - Data Traffic and Growth Rate on Mobile Network

Years Average Data Traffic Annually Annual growth Average annual data traffic/user
(Mb) rate % (Mb)
2013 2,338,745 - 0.53

© IEOM Society International 3077

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

2014 3,358,863 44% 0.84

2015 4,398,942 54% 0.89
2016 6,242,235 85% 1.05
2017 10,221,372 93% 1.50

Fig. 4 - Mobile Traffic Network Data Distribution (2013-2017)

The analysis of the highest and lowest congestion of the Libyan telecommunications company in Tripoli region is
summarized below:
Winter Season; Due to national holidays and celebrations, February saw the highest traffic, contributing to the
increased use of telecommunications and the Internet to transport events online, which will increase traffic this season.
Although traffic is at its lowest in January, it may be because most students are busy for final exams, and not interested
in the Internet.
Spring Season; The highest internet traffic flow was in May. The reason for this is that the big event of the
Telecommunications and Information Technology Exhibition is held in Tripoli every May. While the lowest traffic
occurred in April, an investigation is needed to determine the probable cause.
Summer Season; Students make up a large proportion of internet consumption. In August, most students take summer
vacations, so they have the highest traffic of this month. In addition, the data analysis shows that the lowest internet
traffic occurred in June. However, further investigations are needed to determine the correct probable cause.
Autumn Season: Although the highest Internet traffic was in November and the lowest in September, it still needs to
be examined for the right reasons.

Overall, the internet traffic data analysis shows that August has the highest internet traffic flow in the past five years
(2013-2017) and the lowest traffic in June. Additionally, it should be noted that the average increase in per capita
internet traffic increased gradually between 2013 and 2017, as shown in Figure 4 and Table 1. Internet usage in 2014
and 2015 was significantly slow. In 2017, the network was optimized and upgraded with the introduction of new
HSPA technology. As a result, the average consumption per user jumped from 0.53 in 2013 to 1.5 in 2017. This means
that there will be a strong demand for Internet services in the coming years. As a result, we need to develop a predictive
mobile internet traffic model that can provide predictive capabilities for a variety of purposes, such as network capacity
planning, customized traffic development, quality of service improvement, resource management, and making
decisions that can be used to avoid risks of network congestion.

5.3 Features Selection

In machine learning and statistics, features selection, also known as variables selection to build the model. One
important thing to keep in mind is trade between predictive accuracy and model interpretation. Because if we use a
large number of variables, the predictive accuracy is likely to increase but the interpretation of the model will decrease.
However, there are many ways to choose the model features, some treat it strictly as a work of art, others as a science,
while some form of domain knowledge and a disciplined approach are probably the best solutions. In fact, there is an
excellent method of selecting features, so-called stepwise, which is a combination of two wrapper methods for
selecting by forward and backward selection techniques.

Based on our knowledge of the telecom domain and the mobile internet traffic low historical data, we have selected
three main features of mobile network services data to model internet traffic forecasts as follows: mobile data traffic,
mobile data traffic per user, and mobile users (subscribers). Based on the data from these main features, other variables
were generated as; 5 days moving average data traffic, 10 days moving average (MA) data traffic, 15MA data traffic,
users, 15MA Mobile users, traffic per user at t -1, 5MA consumption per user, 10MA consumption per user, 15MA

© IEOM Society International 3078

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

consumption per user, Data traffic at t -1, Data traffic at t -2, and Data traffic at t -5. Then, stepwise technique was
applied. It shows that the best combination of features is (traffic per use)t -1, users, and (Data traffic)t -1. Therefor, three
inputs are selected to be incorporated into the input layer. The daily observation of mobile traffic data provided from
mobile service during January 01, 2013 to December 31, 2016 was used to build the model and denoted as group A,
while the mobile traffic dataset during 2017 used for evaluating forecast accuracy and denoted as group B. However,
traffic data of group A is randomly divided into training, validation, and testing sets to avoid over-fitting and improve
network model generalization. The training set (68%) is used to update network weights and biases. The validation
set (16%) is used to monitor network performance during the training process. Training can be stopped if the execution
of the model results in a minimum error in the validation dataset. The testing dataset (16%) is used to fine tune the
models. This is not used in training or validation, but to determine the optimal network architecture, select the
appropriate model, and evaluate the performance (generalization) of the fully defined model.

In order to find an efficient network model and analyze the performance, the data must be pre-processed in order to
have a small variation of the output data. It is well known that, in theory, the output data can have a large domain of
variation. However, if the domain of variation of the output data is large, the ANN model tends to be less stable. The
dimensionless output data ranges from 0 to 1, while the input data ranges from −1 to 1. Table 2 shows the results of
the pre-process stage.

Table 2 - Results of The Pre-Process Stage

Parameters (internet traffic)t-1 Users (traffic /user)t-1 internet traffic (Mb)
Column type input Input input Output
Scaling range [-1..1] [-1..1] [-1..1] [0..1]
Min 1042994.76 2316830.3 0.413134 1042994.76
Max 3499321.4 2695701.4 1.353166 3499321.4
Mean 2560315.1 0.835849 2143037.16 2141794.90
Std. deviation 68391.39 0.209051 543270.14 543602.59
Scaling factor 0.000005 2.127586 4.07E-07 8.14E-07

5.4 ANN Model Design

Typically, training a basic multilayer ANN boils down to minimizing some kind of error function. Commonly, the
sum of squared errors is chosen as a function of this error which is often called objective function or cost function or
loss function. The learning problem is presented as a search or optimization problem and an algorithm is used to
navigate the space of possible sets of weights that the model can use to make good or sufficiently good predictions.
To specify the number of hidden layers and the number of units in each layer, a trial and error approach was carried
out, beginning with one hidden layer and one hidden unit. Hidden units were then gradually added. Based on the error
function (Sum of squared error, SSE), several architectures were compared with different network parameters, three
nodes in the input layer and one node in the output layer. hidden layers are between 1 and 3, nodes of hidden layer 1
are between 1 and 10, nodes of hidden layer 2 are between 1 and 6, nodes of hidden layer 3 are between 1 and 4, and
three different types of activation function (AF), called logical, hyperbolic tangent and linear. The nine best
architectures are listed in Table 3. For the best selection, the nine networks were trained with seven different training
algorithms; Quick Propagation, Conjugate Gradient Descent, Quasi-Newton, Limited Memory Quasi-Newton,
Levenberg-Marquardt, Online backpropagation, and Batch back propagation. In each case, the architecture was
retrained seven times with different initial weight randomizations. However, one method to avoid local minima is to
retrain the network with a different weight initialization. This will locate the error function (SSE) at different minima
in the error surface, and then choose the acceptable solution. However, the selected data traffic model is multilayered
with three input nodes and one hidden layer with eight nodes, [3−8−1]. The network was trained using a Quasi-Newton
training algorithm. The neurons in the backpropagation used a logistic function as an input activation function and
linear function as an output activation function. The proposed traffic model provides low minimum errors in training,
validation, testing, and overall dataset, as shown in Table 4. In addition, the statistical error analysis of the prediction
capability obtained from the proposed ANN model is summarized in Table 5. Obviously, the proposed ANN model
provides traffic forecast values with R2 of 97.5% and an average error of 706 Mb, which indicates that the model
describes the data very well.

© IEOM Society International 3079

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Table 3 - The Best Networks Architecture

# Architecture Input AF Output AF SSE Weights Train Error R2,% Test Error
1 [3-8-1] Logistic Logistics 0.000018 41 61562 98.66 55647
2 [3-8-1] Logistic Hyperbolic 0.000018 41 61565 97.35 551000
3 [3-8-1] Logistic Linear 0.000018 26 60872 98.67 55123
4 [3-5-1] Hyperbolic Logistics 0.000018 26 61268 98.66 55324
5 [3-7-1] Hyperbolic Hyperbolic 0.000018 36 61206 98.66 55248
6 [3-3-1] Hyperbolic Linear 0.000018 16 60700 98.66 55041
7 [3-7-1] Linear Logistics 0.000015 36 74346 98.64 68665
8 [3-6-1] Linear Hyperbolic 0.000015 31 74421 98.28 68013
9 [3-4-1] Linear Linear 0.000018 21 62007 98.60 56102

Table 4 – Statistical Error Analysis of Training, Validation, and Testing sets

Mean Mean Relative Min Max

Dataset Std Dev R2,%
Error Error Error Error
Training 60463.8 0.02944 1.423 665461.4 65372.3 93.25
Validation 56263.2 0.027288 553.74 357245.7 50651.5 96.74
Testing 55012.7 0.0262 102.3 346919.1 66622.5 97.85

Table 5 – Statistical Error Analysis of the Proposed ANN Traffic Model

Error Absolute Error Relative Error R2

Dataset
Max Mean Min Max Mean Min Max Mean Min %
Overall 562,263 706 676,618 676,618 60,158 102.3 0.23 0.0014 0.439 97.5

Moreover, Figures 5, 6 and 7 show the scatter plots of actual traffic versus predicted model output for training,
evaluation and overall data points respectively. The Figures show all the points are scattered around the y = x line,
indicating excellent agreement between the actual and estimated data traffic points.

6. Mobile Traffic Forecast for 2017

As mentioned above, daily monitoring of mobile traffic data over the period 2013-2016 (Group A) served to build the
model, while the 2017 Mobile Traffic Data Set (Group B) used for evaluating forecast accuracy.

6.1 Quarterly and Semi-Annual Mobile Traffic forecasts

In the quarterly forecast, the input variables represent the historical data traffic recorded in Q4 2016 and applied to
the model to forecast Q1 2017 mobile traffic. Similarly, for Q2 2017 mobile data traffic forecasts, the inputs are Q1
2017 predicted data traffic. Likewise, in the semi-annual forecast, the input variables represent data traffic recorded
in the second half of 2016 and are applied to the model to forecast mobile data points for the first half of 2017. Table
6 presents the quarterly and semi-annual mobile traffic forecasts. The results of the forecasting error analysis show
that the quarterly forecasting scenarios performed better than the semi-annual forecasting scenarios with the lowest
average error of 54402 Mb and higher R2 of 91%. Other related statistical error analysis is presented in Table 7.
Because the purpose of this study is to develop an accurate model to estimate the demand for data traffic on the Libyan
mobile network. This model enables resource management to allocate optimal capacity resources, better service
quality and cost savings.

© IEOM Society International 3080

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Fig. 5 – Scattered plots of Actual Vs. Model Output traffic; Training Dataset

Fig. 6 – Scattered Plots of Actual Vs. Model Output Traffic; Validation Dataset

Fig. 7 - Scattered Plots of Actual vs. Model Output Traffic; Overall Points of Dataset Group A

Table 6 - Quarterly and Semi-Annual Mobile Traffic Forecasts

Input Variables
` Traffic Forecast Actual Traffic
(internet traffic)t-1 Users (traffic /user)t-1
Q1 2017 2952852 2492787 1.17 2921850 3033873
Q2 2017 3031088 2514026 1.22 3015280 3333114

© IEOM Society International 3081

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Q3 2017 3321784 2475236 1.32 3193825 4204793

Q4 2017 4199030 2499571 1.70 3761733 4502622
Semi-Annual 1 2934892 2552993 1.14 2905240 3184320
Semi-Annual 2 3177239 2503466 1.26 3106358 4353708

Table 7 - Forecasting Performance Error Analysis

Error Absolute Error Relative Error
Period R2,%
Max. Mean Min. Max. Mean Min. Max. Mean Min.
Quarterly 1005347 54402 112022 1005347 544023 112022 0.23 0.133 0.04 91%
Semiannual 1247349 76321 279079 1247349 763214. 279079 0.28 0.18 0.09 62%

7. Conclusions
This study provides an overview of data mining in telecommunications and proposes a novel model for forecasting
mobile internet traffic based on artificial neural networks and draws the following conclusions
• Previous research has shown that telecommunication datasets used in data mining can be classified into; call
details, network and customer data.
• The applications of telecommunications data mining fall into three categories; customer relationship
management, fraud detection, and network management. Most of these studies were devoted to customer
relationship management, such as; customer value and retain profitable customer, generate customer
profitable and mine these for marketing purposes, acquire new customer, maximize profit, and churn analysis
which includes churn prediction and churn management
• Application of ANNs to traffic prediction show that ANN, RBFNN and GMDH-PNN can predict traffic with
very high accuracy.
• Many previous telecommunication data mining research has employed time series forecasting techniques.
Most of these studies did not specify the name of the input features. Only the number of input nodes was
defined.
• Mobile voice and internet data was recorded as a daily traffic monitor from January 1, 2013 to December 31,
2017 on the Libyan real mobile network, which covers the Tripoli region. Between 2013 and 2017, it was
noticed that the demand for voice services increased insignificantly, while the demand for Internet services
increased significantly.
• The trend mining of Internet traffic during the past five years (2013-2017) shows that traffic is highest in
August and lowest in June.
• The daily observation of mobile traffic data provided from mobile service during January 01, 2013 to
December 31, 2016 (group A) was used to build the model, while the dataset during 2017 (group B) used for
evaluating forecast accuracy.
• Data of group A is randomly divided into training, validation, and testing sets to avoid over-fitting and
improve network model generalization. The training set (68%) is used to update network weights and biases.
The validation set (16%) is used to monitor network performance during the training process.
• The proposed ANN forecasting model has three layers with eight nodes in the hidden layer. The network was
trained using a Quasi-Newton training algorithm.
• In features selection, stepwise technique shows that the best combination of features is (traffic per use)t -1,
users, and (Data traffic)t -1
• The neurons in the backpropagation used a logistic function as an input activation function and linear function
as an output activation function.
• The proposed traffic model provides low minimum errors in training, validation, testing, and overall dataset.
It provides traffic forecast values with 97.5% R2 and an average error of 706 Mb, which indicates that the
model describes the data very well.
• The results of the quarterly and semi-annual forecasts of mobile traffic show that the quarterly forecast
performed better than the semi-annual forecast, with its lowest average error of 54402 Mb and higher
correlation coefficient of 91% R2.

References

© IEOM Society International 3082

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Adomavicius, G & Tuzhilin, A. (2001) Expert-driven validation of rule-based user models in personalization
applications. Data Mining and Knowledge Discovery, 5 (2001), pp. 33-58
Agarwal, R. & Ranjan, J. (2013) An empirical investigation on association rule mining in Indian retail industry.
International Journal of Intercultural Information Management 3:3, 226-241.
Akhter,M. Iquebal & Ahamad, M. Gulam (2012). Detecting Telecommunication Fraud using Neural Networks
through Data Mining. International Journal of Scientific & Engineering Research, Volume 3(3).
Aliyu, O. Okpo, U. O Anene E. C. & Abraham, U. U. (2017) Busy hour traffic congestion analysis in mobile
macrocells. Nigerian Journal of Technology (NIJOTECH) Vol. 36, No. 4, pp. 1265–1270.
Al-Rababah, A. & Hani, N. (2004) Component linked based system. In Modern Problems of Radio Engineering,
Telecommunications and Computer Science, 2004. Proceedings of the International Conference. IEEE
Arafat, M. Qusef, A. & Sammour, G. (2019). Detection of Wangiri Telecommunication Fraud Using Ensemble
Learning, IEEE Jordan International Joint Conference on Electrical Engineering and Information Technology
(JEEIT), Amman, Jordan, 2019, pp. 330-335.
Ardhan, S. Satsri, S. Chutchavong V. & Sangaroon, O. (2007) Improved model for traffic fluctuation prediction by
Neural network. International Conference on Control, Automation and Systems, Seoul, Korea, Oct. 17-20
Arumawadu, H.I., Rathnayaka, R.M.K.T., Illangarathne, S.K. (2015) Mining profitability of telecommunication
customers using K-means clustering. J. Data Anal. Inf. Process. 3, 63–71.-
Barabas, M., Boanea, G., Rus, A. B., Dobrota, V., & Domingo-Pascual, J. (2011). Evaluation of network traffic
prediction based on neural networks with multi-task learning and multiresolution decomposition. In 2011 IEEE
7th International Conference on Intelligent Computer Communication and Processing (pp. 95-102). IEEE.
Becker, Richard & Volinsky, Chris & Wilks, Allan. (2010). Fraud Detection in Telecommunications: History and
Lessons Learned. Technometrics. 52. 10.1198/TECH.2009.08136.
Bitrix24 (2019). What is customer churn rate? Bitrix24 Collaboration Platform
<https://www.bitrix24.com/glossary/what-is-customer-churn-rate-definition.php> Accessed December 28, 2019
Blavette, V. (2001). Application of Intelligent Techniques to Telecommunication Fraud Detection,
http://www.eurescom.eu/public/projects/p1000-series/P1007/default.asp, May 2001.
Bolton, R. J. & Hand, D. J. (2002) Statistical fraud detection: A review. Institute of Mathematical Statistics,
17(3):235–255.
Botoca, C., & Pătraşcu, A. (2012, November). Channel quality prediction using neural networks. 10th International
Symposium on Electronics and Telecommunications (pp. 199-202). IEEE.
Camilovic, D (2008) Data mining and CRM in telecommunications. Serb J Manage 3(1):61–72
Chaudhary, P. & Singh, V. (2015). A Data Mining Tool for Network Fault Detection. International Journal of
Computer Science & Communication. 6. 275-280.
Collins, M. (1999 A). Telecommunications crime - part1. Computers and Security, (18):577–586.
Collins, M. (1999 B). Telecommunications crime - part2. Computers and Security, (18):683–692.
Collins, M. (1999 C). Telecommunications crime - part3. Computers and Security, (19):141–148.
Cortez, P. Rio, M. Rocha, M. & Sousa, P. (2006) Internet Traffic Forecasting using Neural Networks. International
Joint Conference on Neural Networks, Vancouver, Canada, July 16-21
Cox, L.A. (2002). Data mining and causal modeling of customer behaviors Telecommunication Systems, 21 (2002),
pp. 349-381
Data Science (2019) < https://www.datascienceblog.net/post/machine-learning/forecasting_vs_prediction/>
Enterprise Integration (2019) 5 Ways Machine Learning keeps track of your network traffic
<https://entint.com/blog/artificial-intelligence/five-ways-machine-learning-keeps-track-of-your-network-
traffic/#gref > Posted on May 23, 2019. Accessed January 07, 2020
Erunkulu, O. O., Nnonye, O. E., Okechukwu, U., & Adewale, A. L. (2019). Prediction of Call Drops in GSM Network
using Artificial Neural Network. Jurnal Teknologi dan Sistem Komputer, 7(1), 38-46.
Estevez, P.A., Held, C.M. & Perez, C.A. (2006) Subscription fraud prevention in telecommunications using fuzzy
rules and neural networks Expert Syst. Appl., 31 (2) (2006), pp. 337-344
Farvaresh, H. & Sepehri, M. M. (2011) A data mining framework for detecting subscription fraud in
telecommunication. Engineering Applications of Artificial Intelligence, Volume 24, (1), Pages 182-194
Farvaresh, Hamid & Sepehri, M. Mehdi. (2011). A data mining framework for detecting subscription fraud in
telecommunication. Engineering Applications of Artificial Intelligence. 24. 182-194.
Fawcett, T. & Provost, F. (1997) Adaptive fraud detection. Data Mining and Knowledge Discovery 1: 291.
https://doi.org/10.1023/A:1009700419189
Fayyad, U., Piatetsky-Shapiro, G. & Smyth, P. (1996). From Data Mining to Knowledge Discovery in Databases. AI
Magazine Volume 17 Number 3.

© IEOM Society International 3083

Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Freeman, E., & Melli, G. (2006). Championing of an LTV model at LTC. SIGKDD Explorations, 8(1), 27-32.
Galbreath, J. & Rogers, T. (1999) Customer relationship leadership: a leadership and motivation model for the twenty-
first century business. The TQM Magazine, Vol. 11 (3), pp.161-171
Global Economic Crime Survey (2016). Adjusting the Lens on Economic Crime.
Https://Www.pwc.com/Gx/En/Economic-Crime-Survey/Pdf/GlobalEconomicCrimeSurvey2016.Pdf.
Gor, R. M. (2003) Industrial statistics and operational management. American Journal of Industrial and Business
Management. ICFAI Business School, Bodakdev, USA
Gowrishankar, S. & Satyanarayana, P.S. (2009) A time series modeling and prediction of wireless network traffic.
International Journal of Interactive Mobile Technologies, vol. 3, no. 1, pp. 53–62.
Halibas, A. S., Matthew, A. C., Pillai, I. G., Reazol, J. H., Delvo, E. G., & Reazol, L. B. (2019). Determining the
Intervening Effects of Exploratory Data Analysis and Feature Engineering in Telecoms Customer Churn
Modelling. In 2019 4th MEC International Conference on Big Data and Smart City (ICBDSC) (pp. 1-7). IEEE.
Haridasan V. & Venkatesh S. (2011) CRM implementation in Indian telecom industry—evaluating the effectiveness
of mobile service providers using data envelopment analysis. Int J Bus Res Manage 2(3):120–127
Hein, Daniel (2019) The Benefits of AI and Machine Learning in Network Monitoring.
<https://solutionsreview.com/network-monitoring/the-benefits-of-ai-and-machine-learning-in-network-
monitoring> Posted on February 13, 2019. Accessed January 07, 2020
Hilas, Constantinos S. (2009) Designing an expert system for fraud detection in private telecommunications networks.
Expert Systems with Applications, Volume 36, Issue 9, November 2009, Pages 11559-11569
Hoath,P. (1998) Telecoms fraud, the gory details. Computer Fraud and Security, 20(1):10–14.
Hung, Shin-Yuan, David C. Yen, and Hsiu-Yu Wang. (2006) Applying data mining to telecom churn management.
Expert Systems with Applications 31.3 (2006): 515-524.
Hwang, H. Jung, T. & Suh, E. (2004). An LTV model and customer segmentation based on customer value: A case
study on the wireless telecommunication industry. Expert Systems with Applications, 26 (2004), pp. 181-188
Johnson, M. (1996). Cause and effect of telecoms fraud. Telecommunication (International Edition), 30(12):80–84.
Joseph, Madhuri V. (2013). Data Mining and Business Intelligence Applications in Telecommunication Industry,
International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-2 (2)
Junsong, W., Jiukun, W., Maohua, Z., & Junjie, W. (2009, June). Prediction of internet traffic based on Elman neural
network. In 2009 Chinese Control and Decision Conference (pp. 1248-1252). IEEE.
Kaski,S. Kangas,J. & Kohonen,T. (1998). Bibliography of self-organizing map (SOM) papers: 1981-1997. In Neural
Computing Surveys, Vol. 1, pages 102–350.
KDD- 89 <https://www.kdnuggets.com/meetings-past/kdd89/index.html> Copyright © 2000 KDnuggets
Khaligh A., Miremadi A. & Aminilari M (2012) The impact of eCRM on loyalty and retention of customers in Iranian
telecommunication sector. Int J Bus Manage 7(2):150–162
Koo, B. Lee, S. Kim, W. & Park, J. H. (2014) Comparative study of short term electric load forecasting. Fifth
International Conference on Intelligent Systems, Modeling and Simulation.
Liu, J., Liu, F. & Ansari, N. (2014) Monitoring and analyzing big traffic data of a large-scale cellular network with
Hadoop. in IEEE Network, vol. 28, no. 4, pp. 32-39, July-August 2014.
Mani, D., Drew, J., Betz, A., & Datta, P (1999). Statistics and data mining techniques for lifetime value modeling.
Fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York.
Markus, E.D. Okereke, O.U. & Agee,J.T. (2012) Predicting Telephone Traffic Congestion using Multi-Layer
Feedforward Neural Networks, Advanced Materials Research Vol. 367 pp 191-198.
Masand, B., Datta, P., Mani, D., & Li, B. (1999). CHAMP: A prototype for automated cellular churn prediction. Data
Mining and Knowledge Discovery, 3(2), 219-225.
Mozer, M.C., Wolniewicz, R., Grimes, D.B. Johnson, E. & Kaushansky H. (2000) Predicting subscriber dissatisfaction
and improving retention in the wireless telecommunications industry, IEEE Transactions on Neural Networks,
11 (2000), pp. 690-696
Oravec, M., Petráš, M., & Pilka, F. (2008). Video traffic prediction using neural networks. Acta Polytechnica
Hungarica, 5(4), 59-78.
Ozovehe, A. (2015) Group Method of Data Handling for Modeling Macrocell Traffic Congestion. Zaria Journal of
Electrical Engineering Technology, Vol. 4 (1&2), March and September, ISSN: 0261-1570, pp. 46 – 52.
Ozovehe, A., Okereke, O. U., Chibuzo, A. E., & Usman, A. U. (2018). Comparative analysis of traffic congestion
prediction models for cellular mobile macrocells. European Journal of Engineering Research and Science, 3(6),
32-38.
Pandey, G., Siddiqui, K. M., & Choudhary, A. K. (2013). Telecom voice traffic prediction for GSM using feed forward
neural network. International Journal of Engineering Science and Technology, 5(3), 505.

© IEOM Society International 3084

 Proceedings of the International Conference on Industrial Engineering and Operations Management
Dubai, UAE, March 10-12, 2020

Pijush, S., Sekhar, S. & Balas, V. E. (2017) Handbook of Neural Computation. London: Academic, an Imprint of
Elsevier, 2017. Print.
Qu, Zhiguo et al (2016). Multilevel pattern mining architecture for automatic network monitoring in heterogeneous
wireless communication networks. In China Communications, vol. 13, no. 7, pp. 108-116.
Raheem, M.A & Okereke, O.U. (2014) A Neural Network Approach to GSM Traffic Congestion Prediction.
American Journal of Engineering Research (AJER) Volume-03, Issue-11, pp-131-138.
Rey, Tim, et al. (2012) Applied Data Mining for Forecasting Using SAS. SAS Institute Copy bibliography citation
Rosset, S., Neumann, E., Eick, U., & Vatnik (2003). Customer lifetime value models for decision support. Data
Mining and Knowledge Discovery, 7(3), 321- 339.
Sanders, H. T. (1998) A contribution to the study of actuarial and individual methods of prediction. Convention center
follies, The Public Interest, 132, 58–72
Sasisekharan, R., V. Seshadri, & S.m. Weiss. (1993) Proactive Network Maintenance Using Machine Learning.
Proceedings of GLOBECOM '93. IEEE Global Telecommunications Conference, doi:10.1109/glocom.
Sasisekharan, R., Seshadri, V., & Weiss, S. M. (1996). Data mining and forecasting in large-scale telecommunication
networks. IEEE Expert, 11(1), 37-43.
Strouse, K. G. (1999) Marketing Telecommunications Services: New Approaches for a Changing Environment,
Artech House, Norwood.
Tripathi, S., Bhardwaj, A., & E, P. (2018). Approaches to Clustering in Customer Segmentation. International Journal
of Engineering & Technology, 7(3.12), 802-807. doi:http://dx.doi.org/10.14419/ijet.v7i3.12.16505
Tuovila, Alicia (2019) <https://www.investopedia.com/terms/f/forecasting.asp> Updated Aug 8, 2019
Velayati, M., Shahriari, M., & Hosseinzadeh Lotfi, F. (2019) Classifying the Customers of Telecommunication
Company in order to Identify Profitable Customers Based on Their First Transaction, Using Decision Tree: A
Case Study of System 780, Journal of Optimization in Industrial Engineering, 12(2).
Wei, C., & Chiu, I (2002). Turning telecommunications call details to churn prediction: A data mining approach.
Expert Systems with Applications, 23(2), 103-112.
Ye, L., Qiuru, C., Haixu, X., Yijun, L., & Guangping, Z. (2013) Customer segmentation for telecom with the k-means
clustering method. Inf. Technol. J. 12(3), 409–413.
Zaleski, A., & Kacprzak, T. (2010). Packet switching networks traffic prediction based on radial basis function neural
network.
Zhao, G. F., Tang, H., Xu, W. B., & Zhang, Y. H. (2004). Application of neural network for traffic forecasting in
telecom networks. In Proceedings of 2004 International Conference on Machine Learning and Cybernetics (IEEE
Cat. No. 04EX826) (Vol. 4, pp. 2607-2611). IEEE.
Zhou, C. & Lin, Z. (2018). Study on fraud detection of telecom industry based on rough set, IEEE 8th Annual
Computing and Communication Workshop and Conference (CCWC), pp. 15-19, 2018, January.

Biographies
Saber Kh. Elmabrouk is a talented veteran with over 20 years of experience teaching and researching in Modeling,
Artificial Intelligence and Data Mining, Big Data, Information System Management, Uncertainty analysis, and Risk
Management. He received his Ph.D. in Petroleum Engineering from Regina University in Canada. Prior to his Ph.D.,
he had earned his Master and Bachelor degrees in Petroleum Engineering from the University of Tripoli, Libya. He is
currently department head of Chemical and Petroleum Engineering at Libyan Academy, Janzour, Libya. In addition,
he is an adjunct faculty with department of Engineering Project Management at Faculty of Engineering, University of
Tripoli, Libya.

Alhasan Nureddin Salem holds a M.Sc. in Engineering Management from Libyan Academy, Janzour in 2019 and a
B.Sc. degree in electrical and electronic engineering from University of Zawia, Libya in 2001. Mr. Alhasan began his
career at ABB in 2002 as an energy management engineer. Subsequently, in 2005, he worked for Libyana Mobile
Phone Company, where he began working as a wireless radio engineer, then supervised the mobile installation phase
of GSM Mobile projects from 2005 to 2007. He is then a mobile switching engineer from 2008 to 2009, and in 2010-
2012 worked as an expert in wireless network quality in the field of performance management. He is a radio network
planning manager in telecommunications projects for the period 2013-2015 and a director of management for 3G and
4G mobile technologies for the period 2016-2017. Currently Mr. Alhassn is an expert in core network strategy and
planning in Libyana Mobile Phone Company.

© IEOM Society International 3085

View publication stats