International Journal of Innovative Technology and Exploring Engineering (IJITEE)

ISSN: 2278-3075, Volume-8 Issue-7 May, 2019

Mitigating Cold Start Problem In A Personalized

Recommender System
Selva Rani B, Ananda Kumar S

 ratings provided by them. Education is one of the most

Abstract: A cold-start problem faced by a recommender system significant domains in which there is an exigency need for
leads to serious causes and ruins the functionality of the entire optimized recommender systems precise to the geographical
system, sometimes responsible for losing new users also due to territories. The period by which a student finishes his/her
poor accuracy in recommendations. Recommendation becomes
very rigid in case of a new recommender system where the product schooling is very crucial and taking the next step towards
details exist but no user started viewing or rating the products yet. his/her higher education defines their entire life. The students
Similarly, when a new product is added the corresponding ratings from remote villages are not conscious about massive amount
are missing or when a new user enters the system, there is lack of of choices for their higher education after schooling with
knowledge about the preferences of the new user. This work nominal fee structure and constraints such as geographical
concentrates on the aforementioned cold-start problems by
designing a hybrid recommender engine for academic choices.
domain, financial status and many more. There exists no
Users’ preferences diverge time to time and domain to domain. well-defined recommender system to carry such choices of
Academia is one such field in which students’ feel more possible and available courses to this vivid generation.
challenging to pick up their course after completing their school, Almost all the recommender systems are constructed to
which determines the future of a student. This may be due to suggest items to users by gathering feedbacks or ratings from
either less perception about the available choices or more
experienced users. Feedbacks are collectedly either explicitly
information overload in the internet. There is no single point of
contact which helps the students to explore and suggest the by prompting the users to rate the items or implicitly by
enormous choices in education. Recommender system is a tool observing the activities of users and their past history of
which suggests the users to find out the best products based on purchases. The two variants of recommender systems are
their tastes and needs. Another bigger challenge in this system is content-based and collaborative filtering. Content-based
missing ratings. Existing user profiles represents the preferences filtering stores the items liked or purchased formerly by the
alone and not the rating about the courses or institutes. This work
proposes such a personalized recommender system which
users and based on his/her interests suggest items or products,
recommends opt courses for a student based on his expected score whereas collaborative filtering identifies users with similar
as well as preference. The proposed methodology was evaluated tastes and suggest items liked or purchased by the other users.
on real data set available from previous year engineering But, both the techniques are in need of huge volume of data
counselling conducted by Anna University. for analysis before suggesting an item to a user, which are
escalating abundant challenging disputes for new
Index Terms: Cold Start, Collaborative Filtering, Knowledge
recommender systems which have no prior ratings or
Base, Personalized Recommendations.
preferences given by users. These sort of cold-start problems
I. INTRODUCTION lead to another variant of recommender systems known as
knowledge-based recommender systems. With the deep
Due to exponential growth of information in the internet, knowledge about the domain, these recommender systems
the usage of the computers also increases and almost every gather users’ preferences explicitly for a better
person depend on recommendation systems in their regular recommendation.
day to day activities to make a better decision [1]. Any In this paper, we aimed at designing a personalized
Recommendation System (RS) offers suggestions on various recommender system for exploring academic choices to the
items like movie, music, holiday plans, hotels, airline students those who are about to finish their schooling. A good
reservation, insurance, books, online courses and many more, recommender system should identify a best match between
either to an individual (personalized RS) or to a group of users items and users interests. In this work, it has been proposed to
(group-based RS). It is a tool which ranks or suggests items to read the preferences from the students and to recommend
its users which may be liked or needed by them [2], [3]. courses based on students’ interest. To achieve this mission, a
E-commerce websites like Flipkart, Uber, Amazon etc. much knowledge base about the institutes and courses offered by the
rely on recommender systems to obtain new customers and to institutes as well as the various streams followed by
retain current users. Hence, in such large electronic networks Tamilnadu state board has been constructed. Initially, the
there exists a rivalry to identify the interests and preferences students are made aware about the various possibilities for
of users towards items based on the feedback as well as higher studies based on their stream of education in schools.
Around 135 groups of stream are offered by Tamilnadu state
Revised Manuscript Received on May 10, 2019 board for higher-secondary
Selva Rani B, School of Information Technology and Engineering,
Vellore Institute of Technology, Vellore, India. education. Based on subjects
Ananda Kumar S, School of Computer Science and Engineering, in each stream, the students
Vellore Institute of Technology, Vellore, India. are eligible for higher studies

Published By:
Blue Eyes Intelligence Engineering
Retrieval Number: G5125058719/19©BEIESP 212 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-7 May, 2019

in various courses and institutes. Fig. 1 shows some of the knowledge-base is not an easy task, since it has to be updated
possible domains which are normally preferred by the periodically so as to avoid incorrect recommendations [5].
students for their higher studies. Knowledge-based systems are further classified as
constraint-based recommendation systems and case-based
recommendation systems [3], [6]–[8]. Based on the similarity
metrics, case-based recommendation systems computes
recommendations [9]–[11]. But constraint-based systems
relate the users’ preferences or interests with the features of
the items defined in knowledge bases [8], [12]. This work
focused on constraint-based recommendation techniques to
recommend courses to students based on predefined
knowledge base, which was usually defined by the set of
variables and set of constraints which are
very much essential for solving the constraint satisfaction
Fig. 1 Various domains for higher studies problems [13], [14]. The required solution will be obtained
only if all the constraint variables are satisfied.
II. RELATED WORKS
III. COLD-START PROBLEM
Recommender systems are intelligent tools designed to
offer personalized services to the users. Some major problems Conventional recommender systems such as collaborative
faced in Recommender systems are quality, sparsity, filtering and content-based filtering deal with the similarity
scalability and first rater [1]. It has also been studied that the measures between the users/items or the users’ personal
most commonly used recommendation systems are designed interest and recommends products/items. But it would to be
as collaborative filtering techniques. Collaborative filtering very tough to do so in a recommender system either for new
methods are classified into memory-based collaborative users or new items since there won’t be any browsing history,
filtering and model-based collaborative filtering. A user’s preferences, past purchase details etc. which is known
collaborative filtering recommender system measures the as a cold start problem in recommender systems. The
similarity between users, predicts an item/product and categories of cold-start problems are system cold-start,
recommends to a target user which is termed as user-based item/product cold-start and user cold-start. This work studied
collaborative filtering. In contrast item-based collaborative how to deal with system cold-start and user cold-start
filtering recommends items/products by computing similarity problems in a recommender system by deploying a hybrid
between items. A content-based filtering system recommends recommender engine.
items to a user based on his/her personal interests [2]. A. System cold-start
Knowledge based filtering is a recommendation technique In case of a new system, knowledge-based recommender
which uses explicit knowledge about users, items and user engine might be helpful to recommend products to the users
preferences and recommends items to users by applying by inferring preferences of the users. Knowledge-based
recommendation criteria [3]. It applies reasoning to recommender systems depend on the features of the items and
recommend which items to which users in which context. It is the knowledge about how users’ interests or preferences are
also known as rule-based recommendation system which met by these features. The required knowledge has been
recommends products to users by referring to decision rules. represented as a set of rules and while receiving users’
Those kinds of systems depend on the creation of knowledge interests or preferences and these rules defines which items
rules which propose items or products to the users which have to be suggested. The target user specifies his/her
coincides the prescription of the rules [4]. Knowledge-based preferences as the item features which are used to construct
recommendation systems are useful in circumstances in which the rules in the knowledge base. Hence, the domain specific
the two traditional recommendation (content-based and knowledge base for the recommender system should be
collaborative-filtering) approaches cannot be applied. populated with the sufficient number of features of items. The
Knowledge-based recommender systems accentuates on item features are mapped with users’ preferences and
explicit knowledge about the domain as well as implicit depending on the similarity measure the recommendation task
knowledge about the users to mine appropriate is initiated.
recommendations. Normally, knowledge-based B. User cold-start
recommendation systems involve a set of constraints and a set Popularity is the common strategy used to overcome the
of products. The constraints are used to describe the products user-cold start problem by recommending popular items to
to be suggested based on the current user desires [2]. The user the users and thereby cutting down the results with the help of
of the system provides his/her preferences in the form of item user’s contextual information. The most popular
specifications and which are designated as constraints items/products are recognized by considering current trends,
internally. Knowledge-based recommender allows a globally popular or a certain special occasions etc. By
malleable mapping between the users’ preferences and collecting auxiliary
product descriptions. It explicitly gathers users’ preferences information about the user
or descriptions about the products and recommends the items such as the device used to
by consulting the knowledge base. Construction of a access, the geographical

Published By:
Blue Eyes Intelligence Engineering
Retrieval Number: G5125058719/19©BEIESP 213 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-7 May, 2019

location of the user and so on the recommender engine would

be able to personalize the product/item to the user very soon.
This work also agonized with user-cold start problem and
hence it had been planned to recommend the courses based on
popularity too.

IV. HYBRID RECOMMENDER

In order to attain enhanced accuracy in recommendation, a
hybrid approach had been adopted. A hybrid recommender
merges more than one recommendation techniques and
produces recommendations. Weighted, mixed and cascaded
are the three important approaches in hybrid systems. This
work concentrates on cascaded technique in which the
prediction of knowledge-based recommender is given as input
to the collaborative filter for further refinement as given in Now the recommender engine initially recommends the
Fig. 2. student about the courses to which they are eligible along with
the list of institutions based on the stream of his/her study at
school as in Fig. 3.

Fig. 2 A Hybrid Recommender for a personalized

recommendation

A. Knowledge-based Recommendation
In a constrained-based recommender system a
knowledge-base is generally constructed using set of
variables and set of constraints [13].
Recommendations are computed according to the Fig. 3 Knowledge-based recommendation – A solution for
knowledge-base and the preferences of the users. system cold-start
Recommendation task identifies the items that meet the These initial recommendations are suggested to the target
expectations of the users. Apart from arts, engineering and student and given as input along with other contextual
medical, the students are not much conscious about numerous information to the collaborative filter to refine the
openings of courses. This work aims at providing better recommendation.
recommendations by collecting the profile of the students B. Collaborative Filtering
explicitly. The profile includes mainly the stream at school, Collaborative filtering (CF) is a way of making
their expected scores in subjects and their domain-preference recommendations based on the similar interests between the
for higher studies. To achieve the objective, a recommender users. This technique operates in user-based or item-based
knowledge base was constructed with all the details about mode. User-based collaborative filtering identifies the group
institutes in Tamilnadu, courses offered by the institutes and of users termed as nearest neighbors with similar tastes to the
target user. Based on the ratings given by the nearest
the eligibility criteria needed for the courses. A sample
neighbors, with a help of a statistical measure a prediction is
knowledge-base with details of students vs, details of institutes
made whether to recommend an item to the target user or not.
vi, required constraints cr and filtering conditions cf for the
Pearson correlation coefficient is one such measure used to
chosen domain academic is presented below: compute the nearest neighbors. If u1 and u2 are two users, R is
Constraint-based Recommender Knowledge-base a matrix which represents the ratings to the items given by the
users and r is the rating given by the user for a particular item.
The similarity between two users is computed by using
Pearson correlation (PC) coefficient as in eq. (1)

Published By:
Blue Eyes Intelligence Engineering
Retrieval Number: G5125058719/19©BEIESP 214 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-7 May, 2019

After finding nearest neighbors the prediction is made to

guess the rating for a new item by the target user using the eq.
(2).

Based on the predicted value, the new item would be

recommended or not be recommended. Item-based
collaborative filtering computes cosine similarity as a classic
metric which is proved to be more accurate. The item features
are represented as vectors and the angle between them
measures the similarity value. The cosine similarity (CS)
between two vectors i and j is found using eq. (3) which may
result values between 0 and 1.
Fig. 4 Particulars of number of students chosen specific branches
at preferred institutes

The similarity measure nearer to 1 represents stronger

similarity between two items. Any two feature vectors x and y
may also be compared with the help of Euclidean distance
(ED) as in eq. (4)

Here, in user-based collaborative filtering instead of rating

to the items, the preferences for the courses given by students
are considered. For instance, by stating his/her expected
cutoff in the forthcoming examination, a target student may
wish to know how many students had chosen B.E. Electronics
and Communication Engineering in the previous year or what
all the branches are obtained by previous year students along
with the college code at which they got admission as
illustrated in Fig. 4. Similarly, in item-based collaborative
filtering the target student may mention his/her community
and provide preferences such as branch he/she wants to study
and the college in which he/she wants to join, by requesting
the required cutoff to obtain admission in the particular course Fig. 5 Particulars of cutoff required to obtain a preferred branch
as in Fig. 5. at preferred institute

V. RESULTS AND DISCUSSION

This proposed system has been implemented in Python 3. A
target student is expected to enter his initial profile with
details such as name, mobile_no, e_mail, district,
stream_of_study and expected_scores in main subjects in the
stream they studied in school. The courses along with institute
names are given as initial recommendations to the students
with the help of knowledge-based recommender. Once the
courses are listed based on constraints, the recommendation
has been refined by using collaborative filtering. The target
student is expected to interact with the system by choosing a
particular course For e.g. Engineering. Then by adding other Fig. 6 Similarity measures vs. Recommendation time
contextual information such as the college_code, community,
branch preference etc. collaborative filtering predicts the
details of courses as per the target student’s expectations.

Published By:
Blue Eyes Intelligence Engineering
Retrieval Number: G5125058719/19©BEIESP 215 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-7 May, 2019

applications,” Int. J. Electron. Commer., vol. 11, no. 2, pp. 11–34, 2006.
13. A. Felfernig, K. Isak, and T. Kruggel, “Testing knowledge-based
recommender applications,” OEGAI Journal, Spec. Issue Recomm.
Syst., vol. 24, no. 4, pp. 12–18, 2005.
14. E. Tsang, Foundations of constraint satisfaction: the classic text.
BoD--Books on Demand, 2014.

AUTHORS PROFILE

Mrs. Selva Rani B is presently associated with

Department of Information Technology as an Assistant
Professor at Vellore Institute of Technology in Vellore,
India. Currently she is working in Recommender
Systems and Multi Criteria Decision Making. She is
Fig. 7 Similarity measures vs. Precision and Recall also interested in Network Security.

VI. CONCLUSION
Dr. Ananda Kumar S is an Associate Professor of
This work aims at predicting the academic options to the School of Computer Science and Engineering at
Vellore Institute of Technology in Vellore, India. His
students and recommending the possibilities too. This work research interests include IoT, Security, Recommender
initially aimed in minimizing the cold start problem and Systems and Sensor Networks. He has published good
sparsity problems that normally is very challenging in number of papers in International conferences and
Journals of repute.
collaborative filtering recommendation systems. The method
demonstrated in this work has eliminated the cold start
problem. The performance of the proposed method was also
evaluated using the standard metrics similar to other
recommendation systems. Three similarity measures were
used and the performance of the proposed method was
assessed. As an extension to this method, feedback can be
collected from the students about their institutions and the
courses so that the performance of this methodology would
still be improved.

REFERENCES
1. E. Vozalis and K. G. Margaritis, “Analysis of recommender systems
algorithms,” in The 6th Hellenic European Conference on Computer
Mathematics & its Applications, 2003, pp. 732–745.
2. A. Felfernig, M. Jeran, G. Ninaus, F. Reinfrank, S. Reiterer, and M.
Stettinger, “Basic approaches in recommendation systems,” in
Recommendation Systems in Software Engineering, Springer, 2014, pp.
15–37.
3. S. Trewin, “Knowledge-based recommender systems,” Encycl. Libr. Inf.
Sci., vol. 69, no. Supplement 32, p. 180, 2000.
4. B. Amini, R. Ibrahim, and M. S. Othman, “Discovering the impact of
knowledge in recommender systems: A comparative study,” arXiv
Prepr. arXiv1109.0166, 2011.
5. M. Zanker, “A collaborative constraint-based meta-level
recommender,” in Proceedings of the 2008 ACM conference on
Recommender systems, 2008, pp. 139–146.
6. N. Mirzadeh, F. Ricci, and M. Bansal, “Feature selection methods for
conversational recommender systems,” in e-Technology, e-Commerce
and e-Service, 2005. EEE’05. Proceedings. The 2005 IEEE
International Conference on, 2005, pp. 772–777.
7. B. Smyth, L. McGinty, J. Reilly, and K. McCarthy, “Compound
critiques for conversational recommender systems,” in Proceedings of
the 2004 IEEE/WIC/ACM International Conference on Web
Intelligence, 2004, pp. 145–151.
8. A. Felfernig, G. Friedrich, D. Jannach, and M. Zanker,
“Constraint-based recommender systems,” in Recommender Systems
Handbook, Springer, 2015, pp. 161–190.
9. D. Bridge, M. H. Göker, L. McGinty, and B. Smyth, “Case-based
recommender systems,” Knowl. Eng. Rev., vol. 20, no. 3, pp. 315–320,
2005.
10. R. Burke, “Knowledge-based recommender systems,” Encycl. Libr. Inf.
Syst., vol. 69, no. Supplement 32, pp. 175–186, 2000.
11. F. Lorenzi and F. Ricci, “Case-based recommender systems: A unifying
view,” in IJCAI Workshop on Intelligent Techniques for Web
Personalization, 2003, pp. 89–113.
12. A. Felfernig, G. Friedrich, D. Jannach, and M. Zanker, “An integrated
environment for the development of knowledge-based recommender

Published By:
Blue Eyes Intelligence Engineering
Retrieval Number: G5125058719/19©BEIESP 216 & Sciences Publication