Predicting Student Exam’s Scores by Analyzing
Social Network Data
Michael Fire, Gilad Katz, Yuval Elovici, Bracha Shapira, and Lior Rokach
Telekom Innovation Laboratories and Information Systems Engineering Department,
Ben-Gurion University of the Negev, Beer-Sheva, Israel
Email: {mickyfi, katzgila, elovici, bshapira, liorrk}@bgu.ac.il
Abstract—In this paper, we propose a novel method for
the prediction of a person’s success in an academic course.
By extracting log data from the course’s website and using
network analysis, we were able to model and visualize the
social interactions among the students in a course. For our
analysis, we extracted a variety of features by using both graph
theory and social networks analysis. Finally, we successfully used
several regression and machine learning techniques to predict the
success of student in a course. An interesting fact uncovered by
this research is that the proposed model has a shown a high
correlation between the grade of a student and that of his “best”
friend.
I. I NTRODUCTION
The ability to predict individual or group success in exams
and courses has been researched in the past four decades [1],
[2]. Accurately predicting students’ exam or course grades
has the potential to help students in various ways; by using
accurate predictions we can detect early on students who have
difficulties with the course materials and help them to improve.
Moreover, using this kind of prediction technique can help in
several other education-related areas [3]:
1) Discriminating among enrolment applicants.
2) Advising students on their majors.
3) Identifying productive programmers.
4) Identifying employees who might profit best from additional training.
5) Improving computer classes for non-CIS majors.
6) Determining the importance of oft-cited predictors of
computer competency, such as gender or math ability.
7) Exploring the relationship between programming abilities and other cognitive reasoning processes.
One common approach for solving this type of prediction
problem is to extract as many attributes as possible, sometimes as many as hundreds. By evaluating the value of each
attribute, researchers can attempt to predict exam grades or
other variables using linear regression or multiple regression
methods. Usually, when using regression, one tries to predict
the dependent variables’ values using independent attributes of
different types. The number of independent variables is very
large and includes age and gender [4], [5], GPA grades [6],
educational level of parents [7], emotional and social factors [8], and even the complexity measure of teacher’s lecture
notes [9]. Dependent variables include course grade [5], [6],
[10], [11], exam grade [8], [12] and even aptitude in computer
programming [13]. Other methods used in tackling the grade
prediction problem are the factor analysis or other classification schemes with statistical analysis [5], [10], [14], [15]. For
example, Rountree et al. [15] used Decision Trees analysis in
order to identify combinations of factors that may assist in
predicting success or failure in a CS1 class. In this paper, we
present a method for building a social network of students
in a course for the purpose of predicting their final exam
grades. The method was tested on the course “Computer and
Network Security”, which was a mandatory course taught by
two of the authors of this paper at Ben-Gurion University. The
social network contains data collected from 185 participating
students from two different departments. The course’s social
network was created by analyzing the implicit and explicit
cooperation among the students while doing their homework
assignments. Using this social network, we extracted various
social attributes, such as the grade of the student’s “best
friend”. We demonstrate - using linear regression - that the
best friend’s grade has a direct influence on a student’s exam
grade. We also demonstrate - by using multiple regression
and machine learning - that other social parameters are also
influential in determining a student’s grade. Moreover, they
can help predict which students are likely to fail the test.
II. R ELATED W ORK
In the past four decades, a considerable amount of research
has gone into detecting factors that affects students’ exam and
course grades. Most of these studies used regression in order
to detect the relevant factors that influence students’ grades.
Petersen and Howe [11] tried to predict grades in introductory
CS courses. They found that previous success in high school
mathematics and science had a positive correlation with the CS
course grade. Mazlack [16] found that a low correlation exists
between the final exam grade in a FORTRAN programming
course and a student’s personal background factors in addition
to a students’ results in an IBM’s Programmer Aptitude Test
(PAT). Konvalina et al. [12] used students’ high school performance and background in mathematics in order to predict final
exam grades in the course “Introduction to Computer Science”.
Butcher and Muth [6] found a linear correlation between
different ACT grades and the final CS course grade. Evans
and Simkin [3] utilized a 100 question survey in order to predict students’ grades. They discovered that the Myers-Briggs
cognitive style [17] can assist in predicting students’ grades.
Henry et al. [18] examined the relationships among attributions
and performance in a computer science course. They found
that people with an optimistic attributional style performed
better in a computer programming course than those with a
pessimistic one. Chamillard [19] used students’ performance in
prior academic courses in order to predict their performance in
a particular course. Bennedsen and Caspersen [8] studied the
influence of emotional and social factors on students’ learning
outcomes in an introductory computer science courses. Surprisingly, they did not find a linear correlation between the
exam grade and the social well-being and emotional health
of the students. Recently, Joseph and Suzanne [2] examined
peer tutoring’s influence on students’ performance in CS1 and
CS2 courses. They found indications that peer tutoring had a
positive impact on a students’ course performances, but did not
have a significant impact on the course final exam grades. In
this paper we too use several regression and machine learning
techniques in order to predict students’ final exam grade in
the course “Computer and Network Security”. In order to carry
out these predictions, we mainly used attributes extracted from
the course’s social network, created by us. Similar techniques
that involve social-network analysis and regression were used
by Lee [20] in analyzing the behavior of del.icio.us users, by
Christakis [21] in researching the spread of obesity, and by
Altshuler et al. in predicting the individual parameters and
social links of smart-phones users [22].
III. M ETHODS AND E XPERIMENTS
To cope with the challenge of predicting students’ grades,
we used regression and machine learning in order to analyze
the social network features which were extracted from the
course social network. We constructed the social network of
the course by using different types of homework assignments.
Some assignments were individual (and done online), while
others required a group effort. After constructing the course’s
social networks, the next step was to extract the relevant features from these networks. We developed a Python code using
the Networkx graph package [23] for this purpose. We then
combined the social network graph features with other student
data collected throughout the course, such as their grades in
the assignments and in the final exam. Finally, we used the
collected features in order to test our hypothesis; a student’s
social network can influence that particular student’s grades.
The rest of this section describes in detail the steps taken in
order to run our experiments and prove our hypothesis.
A. Constructing the Course Social Networks
1) Different Homework Assignments: We constructed the
course social network using the course homework assignments
which were a part of the students’ final course grade. The
assignments consisted of several types:
• Online Assignments: The students received five different
online assignments. These assignments were individual
and every student was expected to solve them on their
own. Every student got a different solution, according
to his private email address. Despite the difference in
•
•
the solution itself, all online assignments could be solved
by using the same techniques. These assignments were
stored in a dedicated website, with every access and every
submission to the website recorded in the site’s log. By
analyzing these logs we were able to identify the students
who had cooperated with each other.
Coding Assignment: Students were given one large
coding assignment to be completed in pairs. The assignment consisted of developing a web proxy with multiple
features.
Theoretical Writing Assignments: The students received two different theoretical writing assignments. The
first assignment was to be done in pairs, while the second
could be done in a group of up to four students.
B. Creating the Class Cooperation Social Network Graphs
By using homework assignments and analyzing the website logs (which contained 10,759 entries), we were able to
construct social networks of explicit and implicit cooperation
among the students. We defined the course’s social network
as a weighted multi-graph G =< V, E >, where V is the
set of vertices in the multi-graph, and each vertex v ∈ V
contains unique information about one of the students that
participated in the course. We defined E as the multi-set
of links in the graph, with each link e ∈ E defined to be
a tuple e = (u, v, t, w) ∈ E, where u, v ∈ V , w is the
weight of the link, and t can be one of the following values:
t ∈ {partner’s links, same computer link, same time link}.
We constructed the cooperation multi-graph links in the following manner. First, we defined the partner’s links to be
the explicit connection among the students who submitted
theoretical or coding assignments together as partners. Secondly, we defined the same computer links to be the implicit
connection among students who used the same computer while
solving the online assignments. In order to discover the same
computer links, we used the online assignments logs to extract
the following data for each user: user name, IP address, and
the browser user-agent string. We used the combination of
user IP address and browser user-agent strings in to fingerprint
computers [24]. By collecting computer fingerprints, we were
able to conclude with a high degree of confidence which users
had worked together on the same computer. We then were
able to define a link between every pair of users that used
the computer. Finally, we defined the same time links to be
the second implicit connection among users who probably
solved the assignments together but submitted their solution
from different computers. To uncover the same time links, we
used the online assignments to extract the timestamp of each
submission. We defined two users as connected to each other
by a time link if the two had accessed our website in almost
the same time on several occasions. For each link, we assigned
a weight according to the number of times such a correlation
between the users was found. For example, if two students
submitted, as partners, three different assignments, then the
weight of the partner’s link between them would be three. In
addition, we also added several information layers to the multi-
graph based on the student’s information. For each vertex,
we added the following information: the student’s department,
assignments grades, and the final exam score. At the end of the
construction phase, a multi-graph with 184 vertices and 360
links was generated. We used the Cytoscape [25] software
tool in order to visualize our multi-graph (see Figure I).
In the visualized graph, different link types were displayed
by using different colors; red for “partner’s link”, yellow
for “same computer link”, and blue for “same time link”.
The strength of each link in the multi-graph was calculated
according to the weight of the link between the vertices
where stronger connections were visualized by bolder lines.
Additionally, the following visualization elements were also
added to the vertices of the multi-graph: a) Each vertex is
represented by a different color according to the student’s
department; blue vertices for students from the Information
Systems Engineering (ISE) department and red for students
from the Computer Science (CS) department. b) Different
vertex sizes were assigned according to the student’s final test
score; the higher the student’s test grade, the bigger the size
of his vertex.
78.5
71.95
81.8
77.6
84.1
77.65
83.4
58.0
70.0
76.2
78.85
71.8
87.55
65.4
90.05
70.3
65.35
89.1
78.75
90.0
73.2
87.35
89.1
81.8
75.9
87.6
83.15
66.45
81.05
83.6
65.05
98.4
59.4
67.15
78.5
79.2
77.6
56.75
89.7
83.9
75.0
70.95
82.5
48.8
89.85
81.45
TABLE I
M ULTI -G RAPH V ISUALIZATION C HARACTERISTICS
Characteristics
Graph Components Number
Total Links
Total Vertices
Total Vertices With Final Grade
ISE Students Vertices
CS Students Vertices
Partner’s Links
Same Computer Links
Same Time Link
Color
Blue
Red
Red
Yellow
Blue
Amount
12
360
184
169
109
75
240
23
96
ipated in the course. This network is a weighted graph where
each link’s strength is the accumulated weight of the links
between a pair of students. Namely, we define
PG‘ =< V, E‘ >,
where E‘ = {(u, v, w‘)|u, v ∈ V and w‘ = ((u,v,t,w)∈E) w}
.
Using the graphs mentioned above, we attempted to estimate
the accuracy of our construction process and whether or not
the topology and attributes of the graphs could be used in
order to predict a student’s final test grade.
1) Calculating the Graph Characteristics: We calculated
several graph characteristics (see Table I): a) general graph
statistics (to be explained later on in this paper), b) the number
of links of each type, and c) the number of graph components.
To evaluate the integrity of our representation of the implicit
and explicit links, we calculated several statistics. For each
link type, we calculated the ratio of links among students
from the same department and links among students from
different departments (inter and intra-departmental links). We
also calculated the percentage of connections among vertices
that were both explicit and implicit.
84.1
97.9
79.7
88.1
85.9
66.45
59.1
53.75
82.15
81.5
C. Predicting Students‘ Final Test Scores
79.7
57.85
60.65
77.6
78.5
83.9
91.6
78.8
84.2
83.25
81.3
68.2
55.6
82.5
86.15
81.05
80.3
85.3
81.5
81.8
68.7
78.7
73.3
84.6
86.7
85.75
77.85
48.45
69.9
69.7
88.8
90.9
80.55
67.1
49.9
76.85
85.3
83.9
43.9
92.3
73.35
66.15
61.1
66.5
60.6
61.65
80.8
88.1
67.85
78.3
70.7
78.7
65.55
72.6
68.5
74.8
63.8
89.0
74.15
89.65
76.7
74.9
95.0
78.95
96.3
76.2
85.3
79.7
74.8
89.4
83.85
79.6
69.45
83.05
48.8
69.1
63.0
77.6
79.7
80.15
74.35
88.35
77.15
81.9
68.35
61.5
89.5
83.05
85.65
72.75
83.8
80.6
79.0
38.25
46.75
86.45
73.15
77.6
88.1
77.85
64.2
97.2
69.6
81.65
92.95
81.35
76.55
67.55
76.5
84.6
76.85
Fig. 1. The course social network multi-graph. Each link color represents
connection of a different type; red is “partner’s link”, yellow is “same
computer link” and blue is “same time link”. The color of each vertex
defines the student’s department and each vertex size is in accordance with
the student’s final test score.
Using the course weighted multi-graph, we also created a
graph describing the social network of the students who partic-
Based on the social network graphs, we developed a Python
code using the Networkx graph package [23]. This code
extracted a feature vector for each student based on the
social network graphs described in the previous section. Using
the extracted features together with the students’ personal
information, such as a student’s department, and assignments
score, we attempted to predict the student’s final test grade
(final-grade). This was done by applying several regression
and machine learning techniques.
1) Features Extraction: Using the students‘ personal information combined with the social network graphs, we extracted
the following features for each student u ∈ V :
Personal Information Features.
• Student assignments scores the score the student received
on each one of his assignments.
• Students department - the department the student belongs
to.
• Student Final Test Grade - (final-test-grade(u))
Topological Features.
• Student’s degree (number of friends) - used to define the
student’s centrality in the network and defined as d(u) :=
|{v|(u, v) ∈ E‘}|
•
The number of friends of u by type - using the multigraph G, we can extract the number friends u has with
respect to a specific connection type:
partners-number(u) :=|{(u, v, t, w) ∈ G|
t = ’partners link’}}|.
same-time-number(u) :=|{(u, v, t, w) ∈ G|
t = ’same time link’}}|.
same-comp-number(u) :=|{(u, v, t, w) ∈ G|
t = ’same computer link’}}|.
• Student Final Test Grade (final-test-grade(u))
Friends Grades Features. We defined several functions and
sets designed to make the friends features definitions clearer
and easier to understand:
• The best friend’s score - we attempted to determine the
influence one’s best friend has on one’s scores. One’s
best friends are defined as the friends with the maximum
connection weight to the student in G‘1
best-friend(u, G‘) := bf(u, G‘) :=
first-element({v|(u, v, w) ∈ G‘ and
w ≥ w‘, ∀(u, v‘, w‘) ∈ G‘}).
•
The second best friend’s score the same as the previous
feature. If a student is u’s best friend after we remove
the first best friend from G‘, then the second best friend
of u is defined to be:
second-best-friend(u, G‘) := sf(u, G‘) :=
best − f riend(u, G‘bf (u) )
•
where Gu :=< V − u, E >.
We define w( u, v), u, v ∈ V to be the weight of edge
(u, v) in the weighted social graph, namely:
w( u, v) := {w|(u, v, w) ∈ E‘}
Using the above definition we can define the following
features:
• The best friend test grade - defined as:
best-friend-grade(u) := final-grade(bf(u))
•
The average and weighted average score of the two best
friends - defined as:
two-best-friends-avg-grade(u) :=
final-grade(bf(u)) + final-grade(sf(u))
2
, and
•
two-best-friends-weighted-avg-grade(u) :=
wu,bf (u) final-grade(bf(u)) + wu,sf (u) final-grade(sf(u))
2
Student’s friends maximum score - defined as
max-fscore(u) := max({final-grade(v)|(u, v) ∈ E‘})
1 If there is more than one “best” friend, we arbitrarily chose one vertex
from the set.
•
Student’s friends minimum score - defined as
min-fscore(u) := min({final-grade(v)|(u, v) ∈ E‘})
•
Student’s average and weighted average friends scores defined as
P
(u,v)∈E final-grade(v)
avg-friends-score(u) :=
d(u)
2) Predicting Grades : Using the R-project statistical software [26], WEKA [27] (a popular suite of machine learning)
and the features defined in the above subsections, we ran
several regression and machine learning algorithms. Our goal
was to find a correlation between the different features and the
students’ final test grade and attempt to predict who among the
students will receive a score of below 60, thus failing the final
test. Using regression, the following two experiments were
run: a) a simple linear-regression in order to find a correlation
between students’ grades and their best friends‘ grades, and b)
a multiple linear regression and step-wise regression in order
to build a module for predicting the students‘ grades. Each
of the modules was evaluated by calculating the regression
P-value and R-square values. We also evaluated different
supervised learning algorithms in an attempt to predict which
students are most likely to fail in the final test. We used
WEKAs C4.5 (J48) decision tree, IBk, NaiveBayes, SMO,
Bagging, AdaBoostM1, RotationForest and RandomForest implementations of the corresponding algorithms. The Bagging,
AdaBoostM1, and RotationForest algorithms were evaluated
using the J48 as the base classifier.
IV. R ESULTS
In the following section, we present the results obtained
using the methods described in the previous section. The
results consist of three parts. First, we present the results of the
statistical test used to validate the integrity of the constructed
social network graph. Secondly, we present the results of
the regression analysis techniques mentioned above. Finally,
we present the results of the machine learning algorithms
mentioned in Section III-C2.
A. Graph Construction Integrity
To assess the integrity of the constructed social networks
graph, we determined how many of the links in the graph are
among students of the same department. The results of the
analysis showed that 99.58% of the explicit “partners links”,
100% of the implicit “same computer links”, and 68.75% of
the implicit “same time links” were among students from the
same department (see Figure 2).
In addition, we calculated how many of our implicit links
(i.e., “same computer link” and “same time link”) were also
explicit links (i.e., “partners links”). The results showed that
52% (12 out of 23) of the implicit “same computer links” and
32% (31 out of 97) of the implicit ”same time links” were
also explicit partners links.
These results are particularly significant considering the
a-priori probabilities. With 109 students in the Information
Systems Engineering department and 75 in the Computer Science department, the a-priori probability of intra-departmental
connection ranges between 0.4 and 0.59 (as one can see,
there are hardly any). Moreover, the a-priori probability of
an implicit link to be explicit as well is 0.007, while more
than half of our implicit links share this trait.
Fig. 3. Linear Regression. Predicting the student’s final test grade using
the “Best-Friend-Grade” feature.
Fig. 2. The distribution of links between students by departments. Each
bar illustrates the link type distribution between inter and intra-departmental
links.
B. Regression Analysis Results
Using the R-project software, we ran several regression
algorithms based on the full features vectors that we extracted
from 163 students, out of which 41 failed the final exam (with
a grade lower than 60). Using the regression algorithms, we
generated and evaluated several prediction models in order to
predict the students’ final grades. These models were mainly
based on their social network attributes.
1) Simple Linear Regression: Using a subset of the social
network features described in Section III-C1, we attempted to
use the simple linear regression model:
y = β0 + β1 x1
Four different features were tested in order to create this simple
regression model: a) Using only the Best-Friend-Grade feature
in our regression model produced a regression model with a
positive slope of 0.2525 (β1 = 0.2525) with R2 (R-square) of
0.0553, mean absolute error (MAE) of 11.08 and a p-value
of 0.002 (see Figure 3), b) Using the Two-Best-Friends-AvgGrade feature produced a regression model with a positive
slope of 0.3219, with R2 of 0.0506, MAE of 11.045, and a
p-value of 0.0038, c) Using the Final exercise grade (Final-ExGrade) feature produced a model with R2 of 0.05458, MAE
of 10.929, and p-value of 0.0027, and d) Using the number of
friends with “same-computer” link type (same-comp-number)
feature produced a model with a slope of -4.708 with R2 of
0.019, and p-value of 0.079.
2) Multiple Regression: In this section, we present the
results obtained using the features presented in Section III-C1
and the general multiple linear regression model:
y = β0 + β1 x1 + βx2 + . . . + βn xn
We constructed a model to predict the final test grades of
the students using different feature combinations. When using
multiple regression algorithms with all the extracted features,
the result was a model that predicted the students’ final test
grades with Multiple R-squared of 0.174, MAE of 10.377 and
p-value of 0.009. When a regression model that uses only the
four exercises grades was created, the result was a p-value
of 0.02 with MAE 10.98. Next, we generated a regression
prediction model using backwards stepwise regression. This
resulted in a worse Multiple R-squared value (with a value of
0.1445), but a slightly better p-value of 0.0001 and with MAE
of 10.644. The model only used five features: the best friend
final grade, first and third exercises grades, friend’s maximum
grade, Final-Ex-Grade, and same computer links number.
3) Machine Learning Results: Using the WEKA software,
we ran several machine learning algorithms in order to predict
which of the 163 students will receive a grade of below 60
and fail the final test. Each of our classifiers was evaluated
using 10-fold cross validation approach. We used the areaunder-curve (AUC) measure in order to evaluate our results.
As expected, the ensemble methods fared best, especially
the Rotation Forest and Adaboost classifiers. When all the
features were used, the Rotation Forest classifier performed
best with an AUC of 0.672. Afterwards, we used T-tests
with significance of 0.05 to compare between the different
classifiers. According to T-test results, the RotationForest
classifiers returned better AUC results than the naive ZeroR
and the simple OneR classifiers.
V. C ONCLUSIONS
To paraphrase an old saying, we believe that, ”If you tell
me what your friend’s grades are, I will tell you what yours
will be”. We attempted to prove this hypothesis using the
social networks of a course taught during the spring semester
of 2010. We constructed the course’s social network graphs
using information we collected from the log of the course
website and by analyzing the course homework assignments.
We tested the integrity of our graphs by comparing the link
distribution of the implicit links to that of the explicit links
and looked for correlation. In addition, we also presented
a visualization of the social network graphs of the course.
Using this visualization, we were able to detect an interesting
phenomenon; when looking closely into the social network
graph, one can see that students‘ final grades are closely
related to those of his friends’ grades (see Figure 1). We
attempted to find an explanation to this phenomenon by using
different regression and machine learning techniques. Using
multiple linear regressions, we were able to prove that a
correlation exists between a students’ final grade and that
of their friends. This correlation has a small p-value, which
strongly supports this hypothesis. Furthermore, the regression
models presented in this paper have R-squared values ranging
from 0.1445 to 0.174, and mean absolute error ranging from
10.337 to 10.98 . These values are not uncommon for human
behavioural studies (this is especially true in the field of
grade prediction, according to Evans and Smikin [3]). We
also discovered interesting phenomenon where their is negative
correlation between the final grade and the same-comp-number
feature (with a slope of -4.708 and p-value of 0.079). This
correlation may indicate that students whom cheat on their
homework assignments and solve their assignments with their
friends on the same computer tend to receive lower final
test grades. Using supervised learning algorithms we created
different classifiers that predicated which students are most
likely fail the test. Our Rotation Forest classifier received an
AUC of 0.672.
We believe this work has two main future research directions. The first is to use different classification methods combined with machine learning algorithms to identify students
who are likely to fail their final exam in other courses. The
second future research direction is to use the social network of
the course combined with text analysis techniques (used on the
homework assignments) to follow the diffusion of information
across the network. We believe it could be very interesting
to examine the correlation between homework plagiarism,
knowledge diffusion patterns, and success in the course’s final
exam.
VI. AVAILABILITY
Anonymous version of the students’ multi-graph social
network topology is available for other researchers to use on
our research group website http://proj.ise.bgu.ac.il/sns/.
R EFERENCES
[1] C. Alspaugh, “Identification of some components of computer programming aptitude,” Journal for Research in Mathematics Education, pp.
89–98, 1972.
[2] J. Cottam, S. Menzel, and J. Greenblatt, “Tutoring for retention,” in
Proceedings of the 42nd ACM technical symposium on Computer science
education. ACM, 2011, pp. 213–218.
[3] G. Evans and M. Simkin, “What best predicts computer proficiency?”
Communications of the ACM, vol. 32, no. 11, pp. 1322–1327, 1989.
[4] R. Deckro and H. Woundenberg, “Mba admission criteria and academic
success,” Decision Sciences, vol. 8, no. 4, pp. 765–769, 1977.
[5] T. Cronan, P. Embry, and S. White, “Identifying factors that influence
performance of non-computing majors in the business computer information systems course,” Journal of Research on Computing in Education,
vol. 21, no. 4, pp. 431–446, 1989.
[6] D. Butcher and W. Muth, “Predicting performance in an introductory
computer science course,” Communications of the ACM, vol. 28, no. 3,
pp. 263–268, 1985.
[7] S. Ting and T. Robinson, “First-year academic success: A prediction
combining cognitive and psychosocial variables for caucasian and
african american students.” Journal of College Student Development,
1998.
[8] J. Bennedsen and M. Caspersen, “Optimists have more fun, but do they
learn better? on the influence of emotional and social factors on learning
introductory computer science,” Computer Science Education, vol. 18,
no. 1, pp. 1–16, 2008.
[9] K. Keen and L. Etzkorn, “Predicting students’ grades in computer
science courses based on complexity measures of teacher’s lecture
notes,” Journal of Computing Sciences in Colleges, vol. 24, no. 5, pp.
44–48, 2009.
[10] G. Fowler and L. Glorfeld, “Predicting aptitude in introductory computing: A classification model.” AEDS Journal, vol. 14, no. 2, pp. 96–109,
1981.
[11] C. Petersen and T. Howe, “Predicting academic success in introduction
to computers.” AEDS Journal, vol. 12, no. 4, pp. 182–91, 1979.
[12] J. Konvalina et al., “Identifying factors influencing computer science
aptitude and achievement.” AEDS Journal, vol. 16, no. 2, pp. 106–12,
1983.
[13] T. Hostetler, “Predicting student success in an introductory programming
course,” ACM SIGCSE Bulletin, vol. 15, no. 3, pp. 40–43, 1983.
[14] P. Campbell and G. McCabe, “Predicting the success of freshmen in a
computer science major,” Communications of the ACM, vol. 27, no. 11,
pp. 1108–1113, 1984.
[15] N. Rountree, J. Rountree, A. Robins, and R. Hannah, “Interacting factors
that predict success and failure in a cs1 course,” in ACM SIGCSE
Bulletin, vol. 36, no. 4. ACM, 2004, pp. 101–104.
[16] L. Mazlack, “Identifying potential to acquire programming skill,” Communications of the ACM, vol. 23, no. 1, pp. 14–17, 1980.
[17] C. Allinson and J. Hayes, “The cognitive style index: A measure of
intuition-analysis for organizational research,” Journal of Management
studies, vol. 33, no. 1, pp. 119–135, 1996.
[18] J. Henry, M. Martinko, and M. Pierce, “Attributional style as a predictor
of success in a first computer science course,” Computers in human
behavior, vol. 9, no. 4, pp. 341–352, 1993.
[19] A. Chamillard, “Using student performance predictions in a computer
science curriculum,” in ACM SIGCSE Bulletin, vol. 38, no. 3. ACM,
2006, pp. 260–264.
[20] K. Lee, “What goes around comes around: an analysis of del. icio. us as
social space,” in Proceedings of the 2006 20th anniversary conference
on Computer supported cooperative work. ACM, 2006, pp. 191–194.
[21] N. Christakis and J. Fowler, “The spread of obesity in a large social
network over 32 years,” New England Journal of Medicine, vol. 357,
no. 4, pp. 370–379, 2007.
[22] Y. Altshuler, N. Aharony, M. Fire, Y. Elovici, and A. Pentland, “Incremental learning with accuracy prediction of social and individual
properties from mobile-phone data,” Arxiv preprint arXiv:1111.4645,
2011.
[23] A. A. Hagberg, D. A. Schult, and P. J. Swart, “Exploring network
structure, dynamics, and function using networkx,” In Proceedings of
the 7th Python in Science Conference (SciPy2008), 2008.
[24] P. Eckersley, “How unique is your web browser?” in Privacy Enhancing
Technologies. Springer, 2010, pp. 1–18.
[25] P. Shannon, A. Markiel, O. Ozier, N. Baliga, J. Wang, D. Ramage,
N. Amin, B. Schwikowski, and T. Ideker, “Cytoscape: a software environment for integrated models of biomolecular interaction networks,”
Genome research, vol. 13, no. 11, pp. 2498–2504, 2003.
[26] R. Team et al., “R: A language and environment for statistical computing,” R Foundation for Statistical Computing Vienna Austria, no. 01/19,
2010.
[27] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H.
Witten, “The weka data mining software: an update,” SIGKDD Explor.
Newsl., vol. 11, pp. 10–18, November 2009. [Online]. Available:
http://doi.acm.org/10.1145/1656274.1656278
Michael Fire