University Recommendation System for Undergraduate Studies in
Bangladesh Using Distributed Machine Learning
Ahmed Nur Merag
a
, Rezwana Chaudhury Raka, Sumya Afroj, Md Humaion Kabir Mehedi
b
and Annajiat Alim Rasel
Department of Computer Science and Engineering, Brac University, 66 Mohakhali, Dhaka 1212, Bangladesh
Keywords:
University Recommendation System, Popularity Based Recommender Model, Collaborative Filtering, KNN,
SVD, Cosine Similarity, Pearson’s Correlation, Data Mining, Distributed Machine Learning.
Abstract:
The study proposes a distributed machine learning-based university recommendation system (URS) in
Bangladesh to help undergraduate students make informed decisions based on user ratings. The system uses
advanced distributed machine learning models such as collaborative filtering and popularity-based recom-
mender model which consists of KNNwithmeans model and singular value decomposition (SVD) model to
process data and provide accurate recommendations, significantly enhancing the university selection process
for students. This study advances educational technology and provides a useful tool for undergraduates in
Bangladesh.
1 INTRODUCTION
Bangladesh, a country famous for its rich cultural
legacy and expanding academic prospects, has seen
an increase in the number of universities and aca-
demic programs catering to a wide range of fields.
With the higher education sector developing, the need
for intelligent and data-driven techniques to assist-
ing students in finding the most appropriate university
and program becomes critical. This research paper in-
cludes a thorough investigation and development of a
URS for undergraduate studies in Bangladesh.
The potential for recommendation systems to
change higher education in Bangladesh is great. Uni-
versities can improve student experiences by uti-
lizing data-driven algorithms, which can help with
course selection, career counseling, student engage-
ment, support services, and alumni networking. This
fusion of technology and education creates a unique,
student-centered environment that supports success
and brings satisfaction. By combining distributed ma-
chine learning and recommendation algorithms, dis-
tributed machine learning highlights this paradigm
and produces scalable, tailored student suggestions.
In order to improve students educational experiences,
this project sets out to build a university recommenda-
a
https://orcid.org/0009-0004-9248-4010
b
https://orcid.org/0000-0002-5759-022X
tion system for undergraduate studies in Bangladesh.
The fundamental goal of this research is to use
two separate recommendation models, the popularity
based recommender model and the collaborative fil-
tering recommender model. Furthermore, the exami-
nation includes advanced strategies such as the KNN
with means model and the SVD - model based collab-
orative filtering. These models work together to pro-
vide a strong framework that uses data analytics and
machine learning to provide university recommenda-
tions to aspiring undergraduates. We have not pro-
vide the actual names of the universities rather used a
unique nameId to identify them discretely.
This study is organized as follows: The earlier re-
search utilizing distributed machine learning and rec-
ommendation algorithms are discussed in section 2.
A short description of the dataset is given in Section
3. Section 4 includes the methodology and section
5 contains performance analysis, accuracy of imple-
mented models and results. Lastly, concluding re-
marks are given in section 6.
2 LITERATURE SURVEY
Li et al. propose a distributed collaborative filter-
ing algorithm using MapReduce for improved per-
formance and scalability in massive recommenda-
352
Merag, A., Raka, R., Afroj, S., Mehedi, M. and Rasel, A.
University Recommendation System for Undergraduate Studies in Bangladesh Using Distributed Machine Learning.
DOI: 10.5220/0012256100003543
In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 1, pages 352-358
ISBN: 978-989-758-670-5; ISSN: 2184-2809
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

tion systems (Li et al., 2014). The authors devel-
oped a distributed matrix factorization method for
recommender systems using Apache Mahout, ad-
dressing challenges in managing collaborative fil-
tering tasks in university recommendation systems
(Ghoting et al., 2011). The researchers showcase
MLlib, a distributed machine learning library for
Apache Spark, for academic institutions’ recommen-
dation systems, benefiting Bangladeshi algorithms
development (Meng et al., 2016). They suggest a dis-
tributed learning framework for personalized recom-
mendation systems, utilizing data segmentation and
parameter synchronization for scalability and com-
munication reduction (Zhang et al., 2019b). The
researchers present federated learning for privacy-
protected university recommendation systems, focus-
ing on design, scalability, and collaborative filtering
(Bonawitz et al., 2019). The authors put forward
a federated collaborative filtering approach for per-
sonalized suggestions, addressing privacy concerns
and demonstrating potential in university recommen-
dation systems(Yang et al., 2020). They introduce
a hybrid recommendation system combining collab-
orative and content-based filtering in a distributed
setting, improving university suggestions quality in
Bangladesh(Jin et al., 2017). The authors recommend
a hybrid distributed recommendation system com-
bining deep learning and collaborative filtering for
improved undergraduate studies in Bangladesh (Yao
et al., 2019). The academics discuss scalability in
distributed machine learning, offering insights for al-
gorithm development. They emphasize the impor-
tance of scalability in recommendation systems for
Bangladeshi institutions, accommodating large user
bases and data volumes (Kumar and Kaur, 2020). The
investigators put forward real-time clustering meth-
ods for Bangladeshi university’s recommender sys-
tems (Islam et al., 2019). The scholars recommend
distributed deep learning methods for massive rec-
ommender systems, focusing on using distributed
computing frameworks to train models, enhancing
functionality and scalability in academic institutions’
recommendation systems (Choi et al., 2018). The
analysts advocate a distributed learning framework
for large-scale heterogeneous collaborative filtering
in their 2019 paper, addressing scalability and data
sources issues in recommendation systems. The paper
offers valuable insights into managing diverse data
in academic recommendation systems (Zhang et al.,
2019a). The investigators develop a distributed rec-
ommendation system using LightFM algorithm for
e-commerce, applicable to university recommenda-
tion systems (Wang et al., 2020). The researchers
explore distributed machine learning for predictive
customization in university recommendation systems,
without direct connection to colleges (Yu et al., 2020).
They analyze hybrid recommendation systems using
collaborative filtering and deep learning, providing
insights for researchers in Bangladesh. This review
study is useful for developing hybrid systems for un-
dergraduate courses (Kaur and Chawla, 2020).
3 DATASET DESCRIPTION
Figure 1: Rating Distribution from 1 to 5.
Figure 2: Rating Distribution-II.
The dataset we have used in this research paper is
created from scratch. The file contains features like
unique userId, university names and ratings collected
from various social media sites which were given out
of 5 (five) stars by the users. Every user is identified
with a unique id here. Furthermore, ratings of 143
universities are given in this dataset as features. From
figure 1, we see that the most frequent rating is 3 and
the least frequent rating is 5. Here, 1 is the lowest rat-
ing and 5 is the highest. In figure 2, we get a clear vi-
sual idea about the density of the ratings. Ratings may
change due to users opinion. In other words, these rat-
ings are optimized to conduct the research smoothly.
University Recommendation System for Undergraduate Studies in Bangladesh Using Distributed Machine Learning
353

4 METHODOLOGY
This research starts with collecting data which is de-
scribed in the dataset description section and then we
move on to data pre-processing and feature engineer-
ing. After that, data analysis and visualization is done.
Consequently, training and testing, model selection
and finally evaluation of the selected models and anal-
ysis of the generated results are performed. From fig-
ure 3, we get a clear idea about the architecture of this
study.
Figure 3: Architectural Schematics.
4.1 Data Pre-Processing
Data pre-processing means cleaning and processing
the collected data. In this step handling missing val-
ues, removing noise, and transforming the data into
a applicable format for the recommendation models
which is incorporated in the URS for undergraduate
studies in bangladesh is done. Also, basic exploratory
data analysis is performed.
4.2 Feature Engineering
Now, we extract important characteristics from the
pre-processed data. This entails encoding categorical
variables, normalizing numerical features, and creat-
ing new features that capture important aspects of the
data.
4.3 Data Analysis and Visualization
The process of obtaining insights and knowledge
from data by utilizing various analytical approaches
and presenting the results in visual formats is referred
to as data analysis and visualization. It means analyz-
ing and interpreting data in order to uncover patterns,
trends, correlations, and other useful information that
can improve decision-making and problem-solving.
We generated rating distribution, top rating count
distribution grouped by the names of universities,
top rating count distribution grouped by the user ids,
mean rating distribution grouped by the names of the
universities, mean rating - rating count distribution
grouped by the names of the universities, mean rat-
ing distribution grouped by user ids and mean rating -
rating count distribution grouped by user Ids.
Figure 4: Top Rating Count Distribution grouped by names.
Firstly, top rating count distribution grouped by
the name of universities graph is generated and shown
below in figure 4.
Figure 5: Top Rating Count Distribution grouped by user
Ids.
Then, in figure 5, top rating count distribution
grouped by user Ids is portrayed.
Figure 6: Mean Rating Distribution grouped by names.
After that, mean rating distribution grouped by
university names is constructed as shown in figure 6.
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
354

Figure 7: Mean Rating - Rating Count Distribution grouped
by names.
Similarly, in figure 7, mean rating - rating count
distribution grouped by the names of the universities
is generated and shown below.
Figure 8: Mean Rating Distribution grouped by User ids.
Subsequently, mean rating distribution grouped by
user ids is formed and shown in figure 8.
Figure 9: Mean Rating - Rating Count Distribution grouped
by User Ids.
Finally, in figure 9, mean rating - rating count dis-
tribution grouped by user Ids graph is shown.
4.4 Training and Testing
We have divided the data into training and test sets in
a 70:30 ratio.
4.5 Model Selection
We have selected 2 (two) appropriate recommenda-
tion models for this study. They are popularity based
recommender model and collaborative filtering rec-
ommender Model. The Collaborative filtering recom-
mender model also includes KNN With means model
and SVD - model based collaborative filtering.
4.6 Popularity Based Recommender
Model
Popularity-based recommender model promotes
products based on popularity in a dataset. Algorithm
replaces popular items with personalized recom-
mendations, assuming interest among large groups
without considering individual choices or activities.
Popularity Score = Number of Interactions (Likes,
Views, etc.) for Item / Total Number of Users
4.7 Collaborative Filtering
Recommender Model
The collaborative filtering recommender model uses
user-item interactions to predict user preferences, en-
abling customized recommendations based on past
agreement and future user interactions.
4.7.1 KNN With Means Model
KNN with Means predicts outcomes by estimat-
ing similarity between users or items, using previ-
ous interactions and calculating K-nearest neighbors
based on individual ratings.When utilizing KNN with
Means, the following equation can be used to forecast
user ’u’s rating of item ’i’:
Figure 10: KNN With Means Formula.
4.7.2 Singular Value Decomposition - SVD
Singular Value Decomposition (SVD) breaks user-
item interaction user’s matrix, singular value,
and item matrices, capturing latent components
and generating recommendations by finding lower-
dimensional representations. The following is the
SVD prediction formula for user ‘u’ and item ‘i’:
Figure 11: SVD Formula.
University Recommendation System for Undergraduate Studies in Bangladesh Using Distributed Machine Learning
355

KNN with Means and SVD are collaborative fil-
tering methods with strengths in local patterns and
global patterns.
5 MODEL EVALUATION AND
RESULT ANALYSIS
At this stage, we take a closer look at the recom-
mendations and assess the performance of the trained
models using appropriate evaluation metrics. This
step helps us to understand how well the models are
performing. We have not published the actual names
of the universities, but rather a unique name Id is used
to identify them discretely. Table 1 is given as an ex-
ample below.
Table 1: Using nameId instead of actual name.
userId nameId University Name
R7S8T9U0 5947683210 BRAC University
Starting with popularity based recommender
model, we try recommendation for three randomly
picked users which are ’A1B2C3D4’, ’E5F6G7H8’
and ’M3N4O5P6’. The university names that are cur-
rently popular are used in the recommendation system
based on popularity. It ranks the names depending
on their popularity, i.e. the number of ratings. If an
institution is highly regarded, It is more likely to be
scored higher and thus recommended. Because it is
dependent on the popularity of the name, it can not be
individualized. Hence, the same list of names will be
recommended to all users.
At first, we have generated the recommendation
table for user ‘A1B2C3D4’ as shown in table 2.
Table 2: Recommendation table for user ‘A1B2C3D4’.
sl num user id nameId score Rank
110 A1B2C3D4 7913620854 113 3.0
141 A1B2C3D4 9826301574 110 7.0
32 A1B2C3D4 3106947285 107 14.0
115 A1B2C3D4 8192574306 107 16.0
18 A1B2C3D4 2356891470 106 19.0
Additionally, a score vs rank graph is also gener-
ated for the user ’A1B2C3D4’ as shown in figure 12.
Figure 12: Score vs Rank graph for user ’A1B2C3D4’.
For user ‘E5F6G7H8’, a recommendation table
which is table 3 and a similar graph which is figure
13 is also generated.
Table 3: Recommendation table for user ’E5F6G7H8’.
sl num user id nameId score Rank
110 E5F6G7H8 7913620854 113 3.0
124 E5F6G7H8 8642097531 112 5.0
137 E5F6G7H8 9617530428 109 12.0
59 E5F6G7H8 4952386170 107 15.0
115 E5F6G7H8 8192574306 107 16.0
Figure 13: Score vs Rank graph for user ’E5F6G7H8’.
Finally, for the third user ‘M3N4O5P6’ a recom-
mendation table is shown in table 4 along with a score
vs rank graph in figure 14.
Table 4: Recommendation table for user ‘M3N4O5P6’.
sl num user id nameId score Rank
28 M3N4O5P6 2937501846 113 2.0
26 M3N4O5P6 2804175963 112 4.0
124 M3N4O5P6 8642097531 112 5.0
137 M3N4O5P6 9617530428 109 12.0
22 M3N4O5P6 2631894750 108 13.0
Figure 14: Score vs Rank graph for user ’M3N4O5P6’.
Because this is a popularity-based recommender
model, we are getting a collection of names that are
nearly identical for all three users. This model has an
evaluation score of 2.84.
Now, we apply collaborative filtering recom-
mender model which consists of KNN With means
and SVD. KNN With means - memory based collab-
orative filtering is done first.
By using user-user collaborative filtering method,
top 5 names are recommended for user ‘A1B2C3D4’
and displayed in table 5.
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
356

Table 5: Recommendation for ‘A1B2C3D4’ using KNN.
sl num userId nameId Rating
0 A1B2C3D4 9876543210 5
1 A1B2C3D4 1234567890 5
2 A1B2C3D4 8192736450 5
3 A1B2C3D4 6512390847 5
4 A1B2C3D4 4728901563 5
Consequently, top 5 names are recommended for
user ‘E5F6G7H8’ and portrayed in table 6.
Table 6: Recommendation for ‘E5F6G7H8’ using KNN.
sl num userId nameId Rating
0 E5F6G7H8 2468135790 5
1 E5F6G7H8 1357924680 5
2 E5F6G7H8 9876543210 5
3 E5F6G7H8 1234567890 5
4 E5F6G7H8 5827369140 5
Lastly, top 5 names are recommended for
‘M3N4O5P6’ which is displayed in table 7.
Table 7: Recommendation for ‘M3N4O5P6’ using KNN.
sl num userId nameId Rating
0 M3N4O5P6 2468135790 5
1 M3N4O5P6 1796253408 5
2 M3N4O5P6 7946832105 5
3 M3N4O5P6 6512390847 5
4 M3N4O5P6 5193640278 5
KNN (K-Nearest Neighbours) With means model
has an test rmse value of 0.0727 and cross validation
rmse value of 0.0843.
Regarding recommendations, each user will have
unique names suggested to them according to the rat-
ings given by comparable users. To locate comparable
users, the KNN model used cosine similarity along
with pearson’s correlation.
Next, we apply the SVD - model based collabo-
rative filtering. Here in table 8, we can see that it
recommended top 5 names for user ‘A1B2C3D4’.
Table 8: Recommendation for ‘A1B2C3D4’ using SVD.
sl num userId nameId Rating
0 A1B2C3D4 5193640278 5
1 A1B2C3D4 7946832105 5
2 A1B2C3D4 1357924680 5
3 A1B2C3D4 8192736450 5
4 A1B2C3D4 6512390847 5
Again, in table 9, recommendation of top 5 names
for user ‘E5F6G7H8’ is shown.
Table 9: Recommendation table ‘E5F6G7H8’ using SVD.
sl num userId nameId Rating
0 E5F6G7H8 5827369140 5
1 E5F6G7H8 4728901563 5
2 E5F6G7H8 5193640278 5
3 E5F6G7H8 5947683210 5
4 E5F6G7H8 8192736450 5
And, top 5 names for user ‘M3N4O5P6’ are gen-
erated and shown below in table 10.
Table 10: Recommendation table ‘M3N4O5P6’ using SVD.
sl num userId nameId Rating
0 M3N4O5P6 1796253408 5
1 M3N4O5P6 8192736450 5
2 M3N4O5P6 7946832105 5
3 M3N4O5P6 3578204961 5
4 M3N4O5P6 6512390847 5
SVD model has a test rmse value of 0.0089 and
cross validation rmse value of 0.0546. Compared to
KNN With means, this model has a lower rmse value.
In terms of suggestions, each user is assigned a
distinct set of names formed by filling in missing en-
tries in the matrix during matrix factorization using
SVD.
Table 11: Model Comparison-I.
Model test rmse cv rmse
SVD 0.0089 0.0546
KNN With Means 0.0727 0.0843
From analysing the results displayed in table 11,
we can now clearly say that SVD is the better model
comparing to KNN with means or popularity-based
system with a superior rmse value of 0.0089.
Some more algorithms or models such as Base-
lineOnly, KNNWithZScore, CoClustering and NMF
are applied on the dataset for a better understanding
and the test-rmse, fit-time and test-time are generated
and exhibited in table 12.
Table 12: Model Comparison-II.
Model test rmse fit time test time
BaselineOnly 0.967806 0.592147 0.375389
KNNWithZScore 1.044329 0.426561 1.312531
CoClustering 1.090479 9.427216 0.254634
NMF 1.132793 17.536839 0.325683
Based on the aforementioned results shown in
both table 11 and 12, we can say that the Baseli-
neOnly model performs well with the data because it
has an rmse score of 0.968 and we can conclude that
in all of the models used in this research, the SVD
model is the best performing model with an rmse
score of 0.0089.
University Recommendation System for Undergraduate Studies in Bangladesh Using Distributed Machine Learning
357

6 CONCLUSION
Recommender systems recommend goods based on
users’ preferences. Popularity-based recommender
models consistently recommend university names for
all users. Collaborative filtering recommender sys-
tems, including KNN With means and SVD, have bet-
ter rmse values and test-results. The SVD model is the
best choice.
To conclude, this study aids Bangladeshi under-
graduates in university selection using multiple user
ratings and distributed machine learning for AI-driven
education solutions.
7 FUTURE WORK
In future, we could work on other universities focused
on different parts of the world.
REFERENCES
Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Inger-
man, A., Ivanov, V., and Wu, X. (2019). Towards fed-
erated learning at scale: System design. arXiv preprint
arXiv:1902.01046.
Choi, J., Kim, J., and Cho, S. (2018). Distributed deep
learning for large-scale recommender systems. In
Proceedings of the 24th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining,
pages 2076–2084. ACM.
Ghoting, A., Talwalkar, A., and Dhillon, I. (2011). Dis-
tributed matrix factorization with mahout. In Proceed-
ings of the 2011 SIAM International Conference on
Data Mining, pages 111–122. Society for Industrial
and Applied Mathematics.
Islam, M. A., Hossain, M. A., and Rahman, M. M. (2019).
Real-time data stream clustering for recommender
systems in big data. IEEE Transactions on Big Data,
5(1):130–143.
Jin, Z., Lu, K., and Liang, S. (2017). A hybrid distributed
recommendation system combining collaborative fil-
tering and content-based filtering. Future Generation
Computer Systems, 75:98–108.
Kaur, H. and Chawla, M. (2020). A hybrid approach of col-
laborative filtering and deep learning for recommen-
dation systems: A comprehensive review. In 2020
IEEE 7th Uttar Pradesh Section International Con-
ference on Electrical, Electronics and Computer En-
gineering (UPCON), pages 1–6. IEEE.
Kumar, N. and Kaur, K. (2020). Scalability in distributed
machine learning. In Distributed Computing and In-
ternet Technology, pages 61–73. Springer.
Li, Y., Wang, S., Zhu, L., Zhang, Y., and Zhang, Z. (2014).
A distributed collaborative filtering algorithm based
on mapreduce for large-scale recommendation sys-
tems. In 2014 IEEE International Conference on Data
Mining Workshops, pages 513–520. IEEE.
Meng, X., Bradley, J., Yavuz, B., Sparks, E., Venkataraman,
S., Liu, D., and Zaharia, M. (2016). Mllib: Machine
learning in apache spark. Journal of Machine Learn-
ing Research, 17(34):1–7.
Wang, H., Xiong, J., Wang, C., Zhang, H., and Shi, Y.
(2020). A distributed recommendation system based
on lightfm in e-commerce. In International Con-
ference on Advanced Data Mining and Applications,
pages 119–130. Springer.
Yang, Q., Liu, Y., Chen, T., and Tong, Y. (2020). Federated
collaborative filtering for privacy-preserving person-
alized recommendation system. Future Generation
Computer Systems, 104:673–682.
Yao, Q., Liu, T., He, X., Wang, J., Gu, X., and Liu, T.
(2019). Hybrid distributed recommendation system
based on collaborative filtering and deep learning. Ap-
plied Sciences, 9(10):2017.
Yu, F., Seo, S., and Brinkhoff, T. (2020). Distributed ma-
chine learning for predictive personalization in intel-
ligent transportation systems. In 2020 IEEE 17th An-
nual Consumer Communications & Networking Con-
ference (CCNC), pages 1–6. IEEE.
Zhang, F., Yuan, N. J., Lian, D., Xie, X., and Ma, W. Y.
(2019a). Distributed learning for large-scale hetero-
geneous collaborative filtering. In Proceedings of
the 25th ACM SIGKDD International Conference on
Knowledge Discovery & Data Mining, pages 1682–
1691. ACM.
Zhang, Y., Liu, M., Zhang, Y., Tang, J., and Gao, H.
(2019b). A distributed learning framework for person-
alized recommendation systems. IEEE Transactions
on Industrial Informatics, 15(3):1471–1481.
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
358