A Review On Recommender Systems For University Admissions

International Journal of Science and Research (IJSR)
ISSN (Online): 2319-7064
Index Copernicus Value (2015): 78.96 | Impact Factor (2015): 6.391
Volume 5 Issue 12, December 2016
www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
A Review on Recommender Systems for University
Admissions
Aditi Bhide1
, Karan Karnik2
, Pradnya Badge3
, Varun Joshi4
, Vina M. Lomte5
1, 2, 3, 4
Department of Computer Engineering, R.M.D. Sinhgad School of Engineering
5
Assistant Professor & Head, Department of Computer Engineering, R.M.D. Sinhgad School of Engineering
Abstract: Data mining is the process of exploring and discovering patterns in large datasets. It involves analyzing a data set from
different aspects and summarizing it into usable and useful information which can be used to minimize costs and increase efficiency.
Students are unaware of the best institute suited for them due to insufficient information. A large number of students apply to
universities for admissions in various streams. This paper explores the various recommender mechanisms that use data mining and
neural networks to tackle university admission problems.
Keywords: data mining; recommender systems; neural networks; university admissions;
1. Introduction
University College education plays an important role in a
student’s life. Hence, choosing the right college for the right
student is a highly challenging task. Although many
universities have adapted to centralized admissions, the
system is not efficient and a lot of time is spent in securing an
admission in the institutes. This is primarily due to large
number of applications being submitted to universities every
year.
Recommender Systems are data filtering mechanisms [1]
which focus on overload of information and filter out
relevant chunks of data according to user preferences. Such
systems have the ability to predict if a user may be interested
in certain items based on their profile [2]. Such systems are
widely used in e-commerce environments where the system
is required to recommend items based on their shopping
history [3].
These models can be applied to university admissions
process as well. Ragab, Mashat and Khedra [4]have proposed
two recommender systems that work together for high
performance. The first system assigns the students a track for
a preparatory year. The second system recommends the
college based on whether the student has passed the first
module or not. The algorithm uses previous historical GPA
scores for predicting the probability of admission.
The rest of the paper is structured as follows: Section 2 gives
an overview about the related work in using recommender
systems for university admissions. Section 3 discusses the
models proposed by Ragab, Mashat and Khedra [4] which
are based on Data Mining and Knowledge Discovery Rules.
It also discusses a model proposed by Simon Fong,
YainWhar Si and Robert P. Biuk-Aghai[] which is based on
Neural Networks and Decision Tree Classifiers. Section 4
compares these models. Finally, the conclusions are
presented in Section 5.
2. Related Work
The system proposed by the authors Simon Fong and Robert
P. Biuk-Aghai uses a composite model of neural networks
and decision tree classification for university admission
recommender system[5]. Every year, thousands of students
apply to different universities. The universities have to
choose the most appropriate students from all the applicants
which can be a very intricate task. Various factors such as
academic scores, gender and caste, proportion of local and
overseas students etc. are taken into consideration while
accepting the applications. In the proposed system the author
suggests that various other factors such as academic
performance can be contemplated from the secondary school
that the students have been to. The students can be broadly
classified into three main subgroups – 1) Top students that
can be recommended for the reputed universities without
any entrance examinations, 2) average students who can
apply for various universities by appearing for their entrance
exams and 3) low performing students. These subgroups can
then be considered at the time of admissions. Thus, using
this system, the universities can select the most suitable
students for the course. Also, students can take factors such
as difficulty, duration, cost etc. of the degree course into
consideration while applying to the university [5].
Greg Linden, Brent Smith and Jeremy York have thoroughly
described the recommendation system used by Amazon.com
[6]. A recommendation algorithm called item-to-item
collaborative filtering is used, which is independent of the
number of customers or the number of products in the item-
catalog. Real-time, high quality recommendations can be
made even on massive data sets using this recommendation
algorithm. In most other recommendation systems, a set of
customers who have purchased and rated the same products
as the user are found out and accordingly recommendations
are made. However, in item-to-item collaborative filtering,
instead of customers, the items that the user has purchased
and rated are identified and similar products are
recommended [6].
An electronic commerce recommendation system proposed
and implemented byShu Liao, Tengyue Zou and Huiyou
Chang[7] which uses association rules and sequential rules
to recommend the appropriate products. Association rules
give information of the products which are often bought
together whereas; sequential rules provide information about
the sequence in which different associated products are
Paper ID: ART20163309 DOI: 10.21275/ART20163309 141

www.ijsr.net
bought. In most of the existing recommendation systems,
only the association rules are considered and validated to
make suggestions, however, in this system the authors have
used both, association rules and sequential rules
simultaneously to get more suitable and appropriate
recommendations.
GopinathGanpathy and K. Arunesh have proposed two new
recommendation models which are Rank-Reco and Test-
Reco based on ranking and testing measures respectively[8].
In the first method, based on ratings provided by the user, a
smart system is constructed. These ratings are then
converted into rankings. The similar aspects are identified in
this method by the numbers of ratings for the items by the
users. In the second method, the inadequacy problem is
taken into consideration.A large number of items are
presented in e-Commerce applications, but sometimes user
only selects and rates afew items. This leads to sparse entries
in the rating vector. In this case the significant user tastes are
predicted based on paired t-test for difference of means. This
two models generate recommendations for web user.
A new concept based on the field of web usage mining, a
new recommender system for user rating and synonyms
based ranking of the websites has been proposed by
Anamika Rajput and Sushil Kumar Chaturvedi[9]. This
paper focuses on data pre-processing which is the first step
of web mining. In this system, when users search for a
keyword then that keyword is searched along with related
synonyms. Websites are displayed based on its search and
user is asked to give rating. By user rating that page is
upgraded or degraded in the database. To differentiate and to
reduce data retrieved map reduce algorithm has been used.
This recommender system is efficient and works faster.
3. Recommender Systems for University
Admissions
A. HRSPCA: Hybrid Recommender System for
Predicting College Admissions
Admission to a university is a crucial factor in determining
the success in carrying through the education. A large
number of students are admissible to universities and the
criteria for the same includes a lot more than just merit.
Factors like student background, gender influence the
eligibility of the students for a particular university.Ragab,
Mashat, Khedra[4] propose an admission system which
makes use of two recommenders to achieve concise and
accurate results. The main goal is to provide the students
useful courses and guidelines in a way that is best suited for
them.The system comprises of two recommenders, for two
different mechanisms, one for recommendation and the other
for prediction. They are cascaded together and applied along
with data mining techniques[10]. The main focus of the
former is to provide guidelines for the students at the
elementary stages Data mining techniques are applied to find
similarities among a huge number of applicant students
data[8]. After the completion of the process the students are
represented with the courses and colleges appropriate for
them while considering their grades and other criteria.
To achieve the desired outcome, the system is divided into
stages all of which are interrelated and only those students
which successfully pass these stages are represented with the
calculated interpretations.
1) Web Portal
The system structure consists of an interactive and
correlative web-portal to ensure communication between
applicants and the system provides facts and data about the
search techniques. The applicant student is provided with a
login ID and a password. The student must enter his/her
personal data which includes name, date of birth, address,
email, mobile number, social status, disability status and
details regarding subsidization. Information about the
school, the merits and grades of the student are registered.
After these preliminary tasks, the available university
courses are provided and selection must be done on the basis
of the desires in priority. A suitable parameter is obtained
from the entered data which can be used for further
calculations.
2) Enrollment stage
The particulars are registered in the system successfully after
the student enters the data. The user id and password is
provided to the student so as to enable access to the system
and continue entering the necessary data in the further
stages.
3) Documentation and auditing
All the necessary certificates and credentials are to be
provided by the students. Data clearing is cultivated in this
stage, i.e. verification, validation, conversion and validation
of the data takes place. This exercise is to verify the
correction authentication of the student’s name, age,
nationality, etc. A mail is sent to the students if any
mismatch in the data is found. After this procedure, the
validated data is stored in the system database. The
documents of the students which are accepted are then sent
further to the track recommender.
4) Track Recommender
The TR is a fundamental track sorting recommender. It
consists of a sorter and a filter. Tracks are made according to
the study tracks available provided by the universities. The
sorter sorts students to these several tracks. After the sorter
finishes its task, the filter is used to re-organize the students
in two classes; the students who pass the courses
successfully are sent to college recommender to be allocated
to certain colleges and the students who fail are dropped and
are postponed of services till they succeed.
5) College Recommender
The recommender contains two components, namely, the
classifier and the allocator. Service is provided to those
students who completed the courses triumphantly. The
classifier categorizes the applicants according to the
qualifications and the other components like the gender,
background etc. the students are then allocated the
respective universities in a just manner [11]. Pattern
discovery rules of data mining are used for classification.
For example, these rules can be used to distinguish the
resident students from the immigrant students. Here a
pseudocode of feature selection and knowledge discovery is
used.

www.ijsr.net
6) College Predictor
The predictor is used as a non-compulsory alternative so as
to provide optional help to students regarding the tracks and
colleges. The current applicant i.e. student data is compared
to the data already available in the system, the historic data.
Thus the students are provided with the most suitable
colleges [12]. The most probable colleges are represented
according to his/her merit and the qualified exam marks in
prep year.
This information provides guidelines to the student to
arrange his desires in priority.
Figure 1: The HRSPCA Architecture
To ensure the smooth functioning of the stages, certain rules
are applied to the TR, CR and certain operations are used by
CP.
7) Rules for TR
The TR focuses on the prep year criteria and the output
allows the students to be clustered in the following five
tracks:
R1: Demographic student information:
The data fields documented in the student personal data are
governed.
R2: Scientific Track:
Different portions of the qualifying exams are considered-
(50% of high school degree) + (30% of aptitude) + (20% of
achievement degree)
R3: Literary Track:
(60% of high school degree) + (40% of aptitude)
R4: Affiliated Track
(50% of high school degree) + (30% of aptitude) + (20% of
achievement degree)
R5: Colleges with no preparatory years:
The colleges which do not have tracks for the prep years are
also considered.
Here, the student data is preprocessed and filtered after the
application of four main rules with the tracks. Furthermore,
gender classification is applied on each rule.
8) Rules for CR:
Data is obtained from the TR and the CR mainly governs
college allocation based on the following criteria:
R1: Passing all prep year courses
R2: Minimum score required for the prerequisite courses
offered by the college
R3: Capacity of the college
Thus various parameters can be established for different
courses, and depending upon the restrictions, the allocation
can be done appropriately.
9) CP Operations:
The qualified students have a high probability of admission
whereas the admission of some of them may depend on
other factors. For example, more students can be admitted if
the college capacity increases. The rest may have no chances
for attending colleges.
Data mining is used in almost all of the stages; associations
and sequences or correlations are used[10]. The data is
therefore represented in a close-packed manner to improve
efficiency. Data clustering and classification are used to
categorize the data present into a format which is domain-
specific[13]. The recommendations are then obtained as the
final output.
B. Hybrid model of Neural Networks and Decision Tree
Classifiers
Fong, Si &Biuk-Aghai [14] believe that university
admissions are dependent upon much more rather than just
test scores. Admissions also rely on student backgrounds
and changing finances and academic interest. Existing
systems depend largely upon yes or no decisions and linear
programming [15].
Student profiles are built by modelling attributes of the
student to the weights of the neural networks. Historical data
is used as a training data set. These networks are adaptive;
they affect the categorization based on change in weights as
and when the system goes through additional training on
recent data [16].
Student data was observed during counselling sessions to
determine the basis on which students choose a university.
Inputs to neural network are decided from these observed
criteria. Each group of students applying to universities has a
distinct neural network system. Hence a huge amount of
computations is required and training the network takes up a
lot of time.
A hybrid model is proposed which takes in a large number
of input records m, and outputs results into n groups. For
admission to universities, the system uses m student profiles
which contain demographic and academic information and n
predicted universities into which a student may get
admission into and the procedure through which the student
should apply.

www.ijsr.net
Figure 2: Classifying Students into different channels
Step1): Back-propagation algorithm [17] is applied to train
and construct a learning model and calculate the importance
of input in correspondence to the output, such as university
recommendation. Sort the input into a list as per feature.
Step2): Input variable is added in the list produced by neural
network to check and validate performance of C4.5
algorithm [18] and error rate is noted. If theerror rate of the
input at iteration i+1 is found to be higher than that of i,
stop. Continue till full list is checked.
Step3): The rules generated by the C4.5algorithm are
checked by measuring its performance. When it satisfies the
criteria set by the user, it is ready to be used with new data.
Selection of correct variables most relevant to admission
into a university is an important factor to achieve high
accuracy of prediction.
In this hybrid recommender model aback-propagation neural
network algorithm was used to sort out the relative
important input variables from the available ones. Once the
neural network identifies the important input variables and
eliminates the rest, the chosen variables are exported to the
C4.5algorithms for generating a decision tree. A
classification tree is then built by the generated rules of
university admissions.
We can find the most favorable values of the weights that
minimize the error between the measured and the output
performance parameters from the neural network model. A
relative importance of input concepts is used to establish a
measure of significance for each input variable by defining
the range of the values between 0 and 1 so that higher values
are associated with more important variables after modeling
in a neural network algorithm.
From the neural network model, we can find the optimum
values of the weights that minimize the error between the
measured and the evaluated (output) performance
parameters. After modeling in a neural network algorithm, a
relative importance of input concepts is used to establish a
measure of significance for each input variable by defining
the range of the values between 0 and 1 so that higher values
are associated with more important variables.
Figure 3: Workflow of RSAU
The loop procedure chooses best inputs for categorization.
1. i = 1;
2. n = number of attributes,
3. attributei represents the rank i attribute in the list
generated by neural networks algorithm
4. Repeat

www.ijsr.net
4.1 select attribute1 to attributei as input attributes to build
classifier by C4.5 algorithm
4.2 calculate error rate ->ei
4.3i++
5. Until i = n or ei> e(i-1)
Output: The optimization classifier model with i attributes.
4. Comparision
The HRSPCA system was implemented using live data and
the web portal was accessed by 66,000 students during the
year 2012 by Khedra et. al. The students who passed the
tests successfully and were allocated standard colleges and
universities were close to 16,000. The students who were
rejected may appear again. The practical experimental
results are shown as follows:
a) TR Outputs:
The student marks are indicated in the output which is
calculated in the sorting process according to the weightage
computed. The weights are calculated by taking the product
of hours and degree. The students are classified using tracks
chosen and the gender is also taken into consideration.
Out of a total of about 19,000 students 28% were
recommended to enter literary track, which included both
male and female students while 58% students for scientific
track. The scientific track included about 2,900 females and
7,000 males. 18% were recommended affiliated track and
only 600 students were recommended no prep year.
b) CR outputs
The capacity of the colleges is a major factor which is
considered while students are to be recommended for all
available university colleges. The outputs include
classification, again, according to gender. The output
includes recommended student’s capacity and the colleges
than can be acquired.
The recommended outputs are compared to the colleges
actually allocated to the students. The system trustability is
achieved and ensured when there is more positive response
from the students.
Neural Networks are being used extensively in the industry
for many years now with adequate library support and
implementations [19]. They are easy to conceptualize. The
performance of the above hybrid algorithm of neural
networks and decision tree classifier was tested with live
student data.
One of the advantages of a decision tree classifier is that it
takes a significantly less amount of time to learn a large
dataset on a medium spec workstation. The proposed
classifier was found to outperform other approaches such as
the C4.5 Classifier and the neural network classifier. Error
rates were found to be significantly higher when the number
of instances was less than 1000, implying that a certain
amount of instances are required to positively affect
performance.
Neural networks using the back-propagation algorithm took
about 20 minutes on the above mentioned platform. The
hybrid model proposed by Fong, Si, Biuk-Aghai [14] uses
back-propagation algorithm for the feature selection and
then applies C4.5 algorithms to generate rules for university
recommendation.
Figure 4: Instance number vs running time comparison of
C4.5 Classifier
Although Neural Networks have a proven track record of
success for certain problem domains, but there are simpler
and more cost effective alternatives available. Training
outcome can be non-deterministic and depend crucially on
the choice of initial parameters. Training a neural work can
take up a large amount of time depending upon the size of a
dataset.
Although the real students’ data used was from Macau, the
design of the recommender is generic and applicable to
educational systems in other countries [20].
5. Conclusion
University admissions are an important part of a student’s
life. Hence, a service must be available that guides the
student properly. The system must be fast, accurate and
light-weight to achieve higher efficiency. The system must
be able to recommend a university based on the student’s
merit and choice. The HRSPCA system allocates students
into suitable streams as well as suitable colleges. The
HRSPCA works on cascaded hybrid recommender based on
knowledge discovery and data mining which leads to
accurate prediction and recommendation.
Neural networks and decision tree classifiers were used to
develop a Recommender System of Admission to
University(RSAU). It analyzes student academic history to
calculate chances of admission to a certain university. The
experiments showed that a hybrid decision tree and neural
network approach improves accuracy in admission to
university classification task and performs substantially
better than a single decision tree orneural network.
Although these systems were tested on Saudi Arabian and
Macau Universities, they are generic and can be applied to
any other universities also.
References
[1] Joseph A. Konstan and John Riedl “Recommender
systems: from algorithms to user experience,” Springer
Science+Business Media B.V. 2012, DOI
10.1007/s11257-011-9112-x.

www.ijsr.net
[2] F.O. Isinkaye, Y.O. Folajimi, B.A. Ojokoh,
“Recommnedation systems: Principles, methods and
evaluation”, Egyptian Informatics Journal (2015) 16,
261–273.
[3] J. Ben Schafer, Joseph Konstan, John Riedl,
“Recommender Systems in E-Commerce,” GroupLens
Research Project.
[4] Abdul Ragab, Abdul Mashat and Ahmed Khedra,
“HRSPCA:Hybrid Recommender System for Predicting
college Admission”,2012 12th International Conference
on Intelligent Systems Design and Applications (ISDA)
2012,107- 113, DOI: 10.1109/ISDA.2012.6416521
[5] Simon Fong, Robert Biuk-Aghai, “An Automated
University Admission Recommender System for
Secondary School Students”, The 6th International
Conference on Information Technology and
Applications (ICITA 2009)
[6] Greg Linden, Brent Smith and Jeremy York,
“Amazon.com Recommendations Item-to-Item
Collaborative Filtering”, Industry Report 2003.
[7] Shu Liao, Tengyue Zou and Huiyou Chang , “An
Association Rules and Sequential Rules Based
Recommendation System”, DOI: 10.1109 / WiCom.
2008 . 2592 ,
Conference: Wireless Communications, Networking
and Mobile Computing, 2008. WiCOM '08.
[8] Anamika Rajput and Sushil Kumar Chaturvedi, “User
Rating Synonyms Based Modified Ranking Technique
for Recommender System”, International Conference on
Computational Intelligence and Communication
Networks, 2015
[9] Gopinath Ganapathy and K.Arunessh, “Models for
Recommender Systems in Web Usage Mining Based on
User Ratings”, World Congress on Engineering VOL I,
July 6 - 8, 2011, London, U.K.
[10]J. W Seifert, “Data Mining: An overview” CRS Report
for Congress, December 16, 2004.
http://www.fas.org/irp/crs/RL31798.pdf
[11]X. Su and T.M. Khoshgoftaar, “A Survey of
Collaborative Filtering Techniques” Advances in
Artificial Intelligence Volune 2009 Article ID: 421425
[12]A. H. M. Ragab, A. F. S Mashat, A. M. Khedra
“Artificial Aided College Recommmender” Paper under
processing, KAU, DSR, Technical Funded Project No.
20-611-D1432, unpublished
[13]D. Mican and N. Tomai “Association Rules based
Recommender System for Personalization in adaptive
Web based Applications.”
http://gplsi.dlsi.ua.es/congresos/qwe10/fitxers/QWE10_
Mican.pdf
[14]Simon Fong, Yain Whar-Si, Robert P. Biuk-Aghai,
“Applying a Hybrid Model of Neural Network and
Decision Tree Classifier for Predicting University
Admission”DOI: 10.1109/ICICS.2009.5397665
Conference: Information, Communications and Signal
Processing, 2009. ICICS 2009.
[15]Gottlieb, E. 2001, “Using integer programming to guide
college admissions decisions: a preliminary report”,
Journal of Computing Sciences in Colleges, Volume 17
, Issue 2, Pages: 271 – 279
[16]Wasserman, P. D. 1989, “Neural Computing”, Van
Nostrand Reinhold, New York.
[17]AP Engelbrecht, I Cloete and JM Zurada.
1995,“Determining the significance of input parameters
using sensitivity analysis. In Proceedings of the
International workshop on Artificial Neural Networks:
From Natural To Artificial Neural Computation (June
07 - 09, 1995)”, Lecture Notes In Computer Science,
Vol. 930. Springer-Verlag, London, Pages: 382-388
[18]Harvinder Chauhan, Anu Chauhan “Implementation of
decision tree algorithm c4.5” International Journal of
Scientific and Research Publications, Volume 3, Issue
10, October 2013 www.ijsrp.org/research-paper-
1013/ijsrp-p2278.pdf
[19]Neural Networks pros and cons
https://www.cs.cmu.edu/~schneide/tut5/node24.html
[20]Alexander, S., Clark, M., Loose, K., Amillo, J., Daniels,
M., Boyle, R., Laxer, C., & Shinners- Kennedy, D.
2003, “Case studies in admissions to and early
performance in computer science degrees”, In
ITiCSEWGR' 03: Working group reports from ITiCSE
on Innovation and technology in computer science
education, Pages:137-147, ACM Press

A Review On Recommender Systems For University Admissions

More Related Content

Similar to A Review On Recommender Systems For University Admissions (20)

More from Becky Gilbert (20)

Recently uploaded (20)

A Review On Recommender Systems For University Admissions