S. Kurbanoğlu et al. (Eds.): IMCW 2010, CCIS 96, pp. 145–153, 2010.
© Springer-Verlag Berlin Heidelberg 2010
Identification of User Patterns in Social Networks by
Data Mining Techniques: Facebook Case
A. Selman Bozkır
1
, S. Güzin Mazman
2
, and Ebru Akçapınar Sezer
1
1
Hacettepe University, Department of Computer Engineering, Ankara, Turkey
selman@cs.hacettepe.edu.tr, ebru@hacettepe.edu.tr
2
Hacettepe University, Department of Computer Education and Instructional Technologies,
Ankara, Turkey
s.guzin@gmail.com
Abstract. Currently, social networks such as Facebook or Twitter are getting
more and more popular due to the opportunities they offer. As of November
2009, Facebook was the most popular and well known social network through-
out the world with over 316 million users. Among the countries, Turkey is in
third place in terms of Facebook users and half of them are younger than 25
years old (students). Turkey has 14 million Facebook members. The success of
Facebook and the rich opportunities offered by social media sites lead to the
creation of new web based applications for social networks and open up new
frontiers. Thus, discovering the usage patterns of social media sites might be
useful in taking decisions about the design and implementation of those applica-
tions as well as educational tools. Therefore, in this study, the factors affecting
“Facebook usage time” and ”Facebook access frequency” are revealed via vari-
ous predictive data mining techniques, based on a questionnaire applied on 570
Facebook users. At the same time, the associations of the students’ opinions on
the contribution of Facebook in an educational aspect are investigated by em-
ploying the association rules method.
Keywords: Social networks, decision trees, Facebook, association rules.
1 Introduction
In recent years, a rapid increase in numbers of social networks along with numbers of
people using these networks has been observed. Social networks, also called social
software or collaborative software, are a range of applications that augment group
interactions and shared spaces for collaboration, social connections, and aggregate
information exchanges in a web-based environment [1]. Similarly, [2] defined social
networks as web-based services allowing individuals to 1) construct a public or semi-
public profile within a bounded system, 2) articulate a list of other users with whom
they share a connection, and (3) view and traverse their list of connections and those
made by others within the system.
Millions of users have been interested in them since the introduction of social net-
work sites (SNSs) such as MySpace, Facebook, Cyworld, Bebo, Twitter, etc. The
majority of these users have integrated such sites into their daily lifes. Because most
146 A.S. Bozkır, S.G. Mazman, and E.A. Sezer
of the social network users are young individuals, many of them are university stu-
dents. Therefore, these sites are considered to play an active role in the younger gen-
eration’s daily life [3], [4]. On the other hand, it has been stated that social networks
have a prominent educational context, and this prominence has prompted a growing
number of educators to consider them to be important sites for student learning al-
though these are not intended primarily as educational applications. Besides, it has
been suggested that these social networks help users re-situate learning in an open-
ended social context by providing opportunities for moving beyond the mere access to
content (learning about) to the social application of knowledge in a constant process
of re-orientation (learning as becoming) [5].
There have been various studies about social networks in the educational context
including using social networks as a tool or utilizing them as an environment for
courses [6], [7], the utility of social networks in the teaching and learning process [8],
their value for communication and collaboration [9], educational usage themes of
social networks (e.g. [10], [11]). However, a study in the literature about data mining
analysis of social network usage has not been encountered.
As one the most popular social networks, Facebook is considered in the present
study. Facebook is defined as “a social utility that helps people share information and
communicate more efficiently with their friends, family and co-workers” (face-
book.com). As of November 2009, with 316 million users, Facebook is the most
popular and well known social network throughout the world. Moreover, Turkey,
with 14 million members, is the third country in terms of number of Facebook users
and half of these members are younger than 25 years old [12].
Data mining is a process that uses a variety of data analysis tools to discover pat-
terns and relations in data that may be used for prediction purposes. Supervised data
mining techniques are used to model an output variable based on one or more input
variables and these models can be used to predict or forecast future cases [13].
The purpose of the present study is to discover some usage patterns (i.e. usage time
and access frequency) of Facebook users by data mining techniques. Additionally, an
attempt is made to reveal the educational associations of the users. It is believed that
social network based application development and educational programs can be en-
hanced by the findings of this study.
2 Data Mining
Data mining is the process of exploration and analysis, by automatic or semi-
automatic means, of large quantities of data in order to discover useful patterns [13].
In other words, data mining is the complete process of revealing useful patterns and
relationships in data by using techniques like artificial intelligence, machine learning
and statistics via advanced data analysis tools. Oracle BI, SPSS Clementine, SAS
Enterprise Miner and Microsoft Analysis Services are well known data mining tools
in the marketplace [14].
Data mining methods are classified into two categories as predictive and descrip-
tive. The aim of predictive methods is to make predictions on unseen cases by using
Identification of User Patterns in Social Networks by Data Mining Techniques 147
seen cases via a trained model. However, the goal of descriptive methods is discover-
ing deep relationships, correlations and descriptive properties of data.
In this study, both of these method groups are employed by using SPSS
Clementine 12. Additionally, various decision trees algorithms such as CART,
CHAID and C5; artificial neural networks (ANN) and SVM (Support Vector
Machine) classifiers in prediction of target variables are used. Furthermore, the
variable importance feature of SPSS Clementine is used in discovering the factors
affecting “Facebook usage” and “Facebook access frequency”. Likewise, the Apriori
algorithm is employed in discovering frequent opinions of students on the educational
benefits of Facebook usage.
2.1 Methodology
As stated previously, various data mining techniques are employed during the analy-
ses and except one (association rules mining discovery), their prediction performances
are compared. Thus, in this section, a brief information is presented about the meth-
odologies followed.
The decision tree method is probably the most popular classification method
among the data mining techniques due to the ease of use and visual interpretation
capabilities. Typically, a data mining task for a decision tree is classification; for
example, to identify the credit risk for each customer [15]. The main idea of a deci-
sion tree is to split the data recursively into subsets so that each subset covers more or
fewer homogeneous states of the dependent variable. At each split in the tree, all
independent variables are recalculated for their impact on the dependent variable.
When this recursive process is stopped and the tree is in a stable state, the required
decision tree is formed [15]. At this stage, new cases can be classified via the deci-
sion tree. This stage is called tree deduction. C5, Quest, CHAID [16] and CART [17]
are well-known decision tree algorithms. Nevertheless, SPSS Clementine serves
whole algorithms in its package. In essence, differentiations among these algorithms
are mainly caused by technical capabilities and employing different splitting ap-
proaches and their functions. For instance, C5 and CHAID algorithms are designed to
classify only discrete valued variables by using “gain ratio” and “gini value” splitting
approaches, respectively. However, CART algorithms are designed for both classifi-
cation and regression purposes.
On the other hand, in the pattern recognition literature, SVM (Support Vector Ma-
chine) is a state-of-the-art method with its powerful discriminative features in linear
and non-linear classifications. Generally, SVM is designed to enlarge the boundary of
any two classes in pattern space by searching for an optimal hyper plane that has
maximum distance to the closest points between two classes which are termed support
vectors [18]. However, SVM has support for multiclass predictions via different de-
veloped kernel functions. By the help of these kernel functions, solving the problems
in upper dimensional spaces becomes possible.
ANN are systems which contain intelligence nodes arranged in layers. In essence,
an ANN has an input layer, a hidden layer, and an output layer. The nodes in the hid-
den layer collect the inputs from the input layer into a single output value which is
148 A.S. Bozkır, S.G. Mazman, and E.A. Sezer
passed on to the output layer. Associated with each node in the network is a weight.
The weights in the network are determined in a training phase of the network using
training data. The network performance is then tested on the remaining data, or hold-
out sample [19].
Association rule mining is again one of the best studied descriptive mining meth-
ods since the first design and creation. Agrawal, Imelinski and Swami stated a new
approach to mining association rules in 1993 and designed a new algorithm, namely
Apriori, via two phases seek mechanism on itemsets and by looking their association
frequencies (Romero & Ventura, 2007). In the second stage of this study, the analyses
are performed by using the algorithm Apriori. In association rules, mining analyzing,
support, rule support, confidence and lift values are the important parameters in the
usefulness evaluation of rules. In this study, lift and support values are considered.
Table 1. Variable names and available answers in the first part of the poll
Variable name Type Available answers and related distributions
Sex Discrete Male (50%) / Female (50%)
Age Discrete 18-25 (74.1%) / 26-35 (20.53%) / 36-40
(3.86%) / 41 and above (1.4%)
Frequency of access to
Facebook
Discrete Once a year (0.18%) / Once a month (2.98%) /
Several times a week (25.26%) / Once a day
(22.81%) / Several times a day (48.77%)
Facebook usage time Discrete Less than 15 mins. (32.28%) / Half an hour
(39.82%) / 1 hour (14.39%) / 1-3 hours (8.6%)
/ More than 3 hours (4.74%)
Education level
Membership in any group
Membership in student
groups
Membership in common
interest groups
Membership in internet &
tech groups
Membership in
organizations
Discrete
Discrete
Discrete
Discrete
Discrete
Discrete
High School (5.96%) / Bachelor (70.35%) /
Master (23.16%)
Yes (99.82%) / No (0.18%)
Yes (86.49%) / No (13.51%)
Yes (77.54.5) / No (22.46%)
Yes (27.02%) / No (72.98%)
Yes (61.93%) / No (38.07%)
3 Data
Data was collected from 570 active Turkish Facebook users (students) with an online
poll. This online poll consisted of two sections. In the first section, demographic
characteristics of Facebook users and their frequency of Facebook usage, length of
time spent on Facebook, and memberships in Facebook groups were collected. In the
second section, a 10-point Likert scale with 11 opinions were asked, the answers
ranging from 1 (strongly disagree) to 10 (strongly agree), like “Facebook contributes
to communication between classmates”, “It’s useful for assigning tasks in classes and
Identification of User Patterns in Social Networks by Data Mining Techniques 149
homework assignments”. Thus members’ views of Facebook in relation to its educa-
tional usage were sought.
The variable names of the first part and available answers are given in Table 1.
Although the initial dataset size was larger than 570 people, during the data cleaning
and transforming steps, 13 people were removed due to the absence of sufficient in-
formation. Therefore, the final dataset comprised 570 people. In the dataset, male and
female participants are almost equal and more than 400 applicants are in the 18-25
age range. Furthermore, almost all students are at either undergraduate or graduate
level.
4 Application of Data Mining
To discover important factors that affect Facebook usage time and access frequency
to Facebook, CART, CHAID, C5, artificial neural network and SVM algorithms,
which are built in to SPSS Clementine 12, were employed on the dataset at hand (see
Fig. 1). The overall data is partitioned as 80% training and 20% testing, respectively.
Training and test datasets are selected randomly. As the dataset consists of discrete
valued variables, the true and false prediction rates are listed.
According to the results (see Table 2), SVM achieves the most accurate predictions
for two target variables. Therefore, it is considered that the variable importance re-
sults of SVM are the most accurate predictions. As can be seen in Fig. 2, sex, educa-
tion level, membership in a group and membership in any common interest groups are
the most important factors affecting Facebook usage time. Sex plays a crucial role in
Facebook usage time with 68%. Again, it can be clearly seen that age, membership in
student groups and usage time variables are the most important factors affecting
access frequency to Facebook. The effect of age is more than 80% in access
frequency.
Table 2. Applied algorithms and prediction results
Target variable - Applied algorithm True classification False classification
Facebook usage – SVM 62.63 % 37.37 %
Facebook usage – ANN
Facebook usage – C5
47.72 %
47.54 %
52.28 %
52.46 %
Facebook usage – CART 43.68 % 56.32 %
Facebook usage – CHAID
Access frequency to Facebook – SVM
41.40 %
69.65 %
58.60 %
30.35 %
Access frequency to Facebook – C5
Access frequency to Facebook – CART
Access frequency to Facebook – CHAID
Access frequency to Facebook – ANN
55.79 %
52.81 %
50.35 %
48.77 %
44.21 %
47.19 %
49.65 %
51.23 %