A Decision Support System For Evaluating International Student Applications

A Decision Support System for Evaluating International Student Applications
Peter Haddawy and Nguyen Thi Ngoc Hien

{[email protected], [email protected]}
Computer Science and Information Management Program
Asian Institute of Technology
1. Introduction
The Asian Institute of Technology is an information on past and current
applicants. We decided to make use of
autonomous postgraduate institution
with a highly international student body. this information by applying data mining
The institute has an enrollment of 2000 techniques to predict student
students from 45 different countries with performance at AIT based on the
the majority nationality comprising only information contained in the application.
35% of the total student body. Each Decision support systems have been
year we receive applications from built to help advisors instruct students in
students who have completed previous choosing suitable courses and
bachelors or masters degrees at any one appropriate study plans [1,2]. Previous
of approximately 600 different work on student performance prediction
institutions. This extreme diversity in has used logistic regression to examine
applicants for admission makes accurate the impact of various factors on student
evaluation a highly challenging task. performance [1]. Bekele and Menzel [5]
The evaluation of applications has been used Bayesian networks to predict
traditionally performed by faculty mathematics performance of high school
members who have some degree of students. Their model categorized
knowledge each particular country’s students into three categories: below
educational system. Unfortunately, it is satisfactory, satisfactory, and above
not always possible to find a faculty satisfactory. The work reported in the
member in each program who has good present paper differs from theirs in
knowledge of each country’s system of highly international nature of the
higher education. We were thus applicant pool and the more fine grained
motivated to produce a decision support prediction.
system to provide a more methodical The rest of this paper is organized as
approach. follows. Section 2 provides background
Research on academic performance [3,4] on Bayesian networks and the
suggests using student outcome as a methodology for their application to this
good basis to assess applicants’ problem. Section 3 presents empirical
qualifications. A performance prediction evaluation of the prediction model. The
model can be built by applying data paper ends with conclusions and a
mining to available admission and discussion of ongoing and future work.
graduation grade point average data.
Fortunately, AIT has a large database of
2. Methodology program (master or doctor) to which the
A Bayesian network [6] is a graphical applicant is applying. The attributes
representation of a probability major and field of study have a large
distribution. It is a directed acyclic number of values so they were processed
graph in which nodes represent random by clustering the values into groups of
variables and links represent similar majors and similar fields of
probabilistic influences between the study.
variables. Probabilistic dependence and The attributes nationality and institute of
independence are expressed by the the previous degree have large number
presence or lack of paths between nodes of values (86 and 1707, respectively)
in the graph. The fact that probabilistic without any intrinsic meaning. We thus
dependence is encoded in the network decided to transform them into more
topology in this way permits probability meaningful values.
distributions over large numbers of
random variables to be compactly The socio-economic environment can
represented and permits calculations to play a major role in the performance of
be performed efficiently. Due to the students. So we used the World Bank
inherent uncertainty of the performance classification of countries according to
prediction problem, we chose to use their Gross National Income (GNI) to
Bayesian networks for the modeling group countries into GNI categories of
task. Using a probabilistic model has the LIC (Low Income), LMC (Low Middle
advantage that it can later become a Income), NOC (High Income, non-
component of a higher level OECD), OEC (High Income, OECD)
optimization model. and UMC (Upper Middle Income).
At AIT coursework is generally The most important factor concerning
completed within the first year of study. the university of the previous degree is
So we used the grade point average the quality of the academic programs
(GPA) accumulated after the first year as there. We gauged this by correlating the
the dependent variable. The numeric GPA of the previous degree with that
GPA at AIT ranges from 0 to 4.0, which obtained at AIT. If students consistently
is also translated into a letter grade with enter AIT with somewhat average grades
values of A, B+, B, C+ and Fail (C or from a particular university but graduate
below). The number of students with high grades, we take this as
classified as Fail is low due to an evidence of the high quality of that
institute policy of continuous review of institution. We rated institutions on a
students and either special tutoring for or scale from 0 to 10.
dismissal of students with low grades. Experimentation with a number of
Based on research on student different Bayesian network structures
performance [3, 5] and available showed that the simple Naïve Bayes
attributes in the admission data, ten model shown in Figure 2.2 produced the
attributes were chosen as predictors of best results. Predictor attributes are
performance: age, gender, marital status, displayed along the sides and top of the
nationality, English test score, institute graph. The model contains two nodes
of the previous degree, major of for the predicted variable CGPA. The
previous degree, GPA of the previous one from which the edges to the other
degree, field of study and degree attributes emanate is a discrete valued
node and the other with the double edge n
is continuous. The continuous CGPA å| a - p |

i =1
i i
represents a set of conditional Gaussian MAE =

n
distributions. This was done in order to
be able to provide a numeric CGPA where ai is the actual CGPA, pi is the
prediction under the limitation that predicted CGPA, and n is the number of
Bayesian networks do not permit data points. The average mean absolute
continuous variables to be parents of error over the three folds was 0.267, or
discrete variables. The model was built about a quarter grade point.
using the Hugin Researcher 6.5
software. When the values of the Actual Predicted Class
predictor variables are specified the Class A B+ B C+ Fail
model provides a probability distribution A 124 240 22 1 0 387
over the predicted letter grades and a B+ 111 612 87 3 0 813
numeric GPA prediction. B 24 251 127 7 0 409
),(/' RI '(*5((
C+ 5 25 35 2 1 68
678'<
Fail 0 9 1 1 0 11
$*( ( 1 * / B6&25( 264 1137 272 14 1 1688
*(1'(5 35( B, 1 6 7 , 7 8 7 ( Table 3.1: Confusion matrix
&*3$
0$5,7$/ 35( B&*3$
*1, 35( B0 $ - 2 5 4. Conclusions and future work

&B*3$ In today’s transnational admissions
Figure 2.2: Naïve Bayes model for environment, educational institutions are
performance prediction. facing the need for a precise and
thoughtful method to evaluate and select
the most qualified applicants graduating
3. Evaluation and results from various institutions in many
To train and test the model we used countries. Our approach using Bayesian
admissions data from 2003 through 2006 networks for predicting applicant
consisting of 1386 masters and 302 performance has the potential to meet
doctoral students. In the data, the CGPA this need. The technique can be applied
had mean value of 3.43 and variance of at any institution that has a good
0.166. Stratified three-fold cross- database of student and applicant
validation was used to evaluate the information.
model. The data was separated so that We are currently working to increase the
each fold had the same distribution of accuracy of the model using some ideas
grades as the entire data set. To predict from ensemble classifiers. We are also
the letter grade, we used the grade with adding a component that will
the maximum probability. The supplement the prediction of student
confusion matrix showing the predicted performance by displaying the past
versus actual result is shown in Table student who is most similar to the
3.1. We evaluated the numeric CGPA applicant being evaluated. The reason
prediction using the mean absolute error, for doing this is to increase user
which is defined as confidence in the system’s predictions.
We are implementing this using case- A Case with Ethiopian Students.
based reasoning, which retrieves past Proc. IASTED International
cases that are similar to the current Conference on Artificial Intelligence
problem being considered. The and Applications , 2005.
challenge here is to define similarity of 6. Finn Jensen, Bayesian Networks and
cases in a way that is consistent with the Decision Graphs, Springer-Verlag,
prediction model. Our approach is to 2002.
use the prediction model itself to
compute the similarity. Finally, we are
working to incorporate the performance
prediction system into an online
admissions processing system being
developed so that the prediction will be
displayed alongside the applicant
information at the time of evaluation for
admission. This will tie into the
institute’s new online application system
(www.ait.ac.th/apply), which permits
students to submit all application
materials over the web.
REFERENCES
1. Ashraful A. Chowdhury, Predicting
success of a beginning computer
course using logistic regression .
ACM conference on Computer
Science, 1987.
2. Alex Dekhytar & Judy Goldsmith,
The Bayesian advisor project .
Available online:
http://www.cs.engr.uky.edu/~golds
mit/papers/#BAP
3. Iida Hakkinen, Do University
entrance exams predict a cademic
achievement? Working Paper Series,
Department of Economics, Uppsala
University, 2004.
4. Paul Golding & Opal Donaldson,
Predicting academic performance.
Proc. 36th ASEE/IEEE Frontiers in
Education Conference , 2006.
5. Rahel Bekele & Wolfgang Menzel,
A Bayesian approach to predict
performance of a student (BAPPS):

A Decision Support System For Evaluating International Student Applications

Uploaded by

Copyright:

Available Formats

A Decision Support System For Evaluating International Student Applications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Decision Support System For Evaluating International Student Applications

Uploaded by

Copyright:

Available Formats

A Decision Support System for Evaluating International Student Applications

Peter Haddawy and Nguyen Thi Ngoc Hien

is continuous. The continuous CGPA å| a - p |

represents a set of conditional Gaussian MAE =

*1, 35( B0 $ - 2 5 4. Conclusions and future work

You might also like