See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342987712

An Enhanced Collaborative Filtering-based Approach for Recommender

Systems

Article in International Journal of Computer Applications · July 2020

DOI: 10.5120/ijca2020920531

CITATIONS READS
12 1,420

3 authors:

Rouhia Sallam Mahmoud Hussein

Menoufia University Menoufia University
6 PUBLICATIONS 52 CITATIONS 57 PUBLICATIONS 408 CITATIONS

SEE PROFILE SEE PROFILE

Hamdy M. Mousa
Menoufia University - Faculty of Computers and Information
62 PUBLICATIONS 305 CITATIONS

SEE PROFILE

All content following this page was uploaded by Rouhia Sallam on 16 July 2020.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

An Enhanced Collaborative Filtering-based Approach for

Recommender Systems

Rouhia M. Sallam Mahmoud Hussein Hamdy M. Mousa

Faculty of Computers and Faculty of Computers and Faculty of Computers and
Information Technology Information Information
Taiz University Menoufia University Menoufia University
Yemen Egypt Egypt

ABSTRACT similarity between users or items. They are divided into two
Recommender systems are software applications that provide categories: user-based and item-based CF [2]. User-based CF
product recommendations for users based on their purchase identifies similar users to the target user for whom the rating
history or ratings of items. The product recommendations are predictions are being computed. In item-based CF the
likely to be of interest to the users and encompass items such as similarities need to be computed between items rather than
books, music CDs, movies, restaurants, documents (news users. In the model-based approaches, machine learning and
articles, medical texts, and Wikipedia articles), and other data mining methods are used for predictive models [2].
services. In this paper, we propose a framework for Model-based approaches use the ratings to learn a model in
collaborative filtering to enhance recommendation accuracy. order to improve the performance of CF. Examples of these
The proposed approach summarized in two steps: (1) item- techniques include dimensionality reduction techniques such as
based collaborative filtering and (2) singular-value- singular value decomposition (SVD), the matrix completion
decomposition-based collaborative filtering. In item-based technique, latent semantic methods, and regression and
collaborative filtering, the similarity between the target item clustering [1] .
and any other item is calculated. Then, the most similar items In the proposed approach, we have combined the best methods
are recommended. The Singular Value Decomposition based in collaborative filtering. Item-based CF provides better
approach handles the problem of scalability and sparsity posed performance. The SVD-based approach handles the problem of
by collaborative filtering and improves the performance of scalability and sparsity in CF and improves the performance of
item-based collaborative filtering. We have tested the proposed recommender systems.
approach by A Large-Scale Arabic Book Reviews (LABR)
dataset. We used four different datasets to compare our In this paper, we analyze the user-item matrix to identify
approach with existing work. The proposed approach evaluated relationships between different items of item-based CF. Then
using the most common metrics found in the collaborative we use these relationships to indirectly compute
filtering: the mean absolute error (MAE) and the root mean recommendations for users [3].The item-based approach has
squared error (RMSE). The proposed approach achieved high two key processes: (a) computing the similarity between each
performance and obtained minimum errors in terms of RMSE pair of items using various similarity measures like cosine and
and MAE values. Pearson metrics [2] and (b) computing the prediction. Item-
based CF provides better performance and quality than user-
Keywords based algorithms in most published research [3, 4,5].
Collaborative filtering (CF), k-Nearest Neighbors (KNN),
This work uses the model-based technique by applying the
Item-based collaborative, filtering Matrix Factorization (MF)
matrix factorization algorithm via SVD. Matrix-factorization-
Singular Value Decomposition (SVD), the mean absolute error
based CF aims to reduce the dimensions of the rating matrix
(MAE), root mean squared error (RMSE).
and discover potential features under the rating matrix for
1.INTRODUCTION recommendation [6, 7].
Recommender systems are one of the important techniques in The SVD technique produces high-quality recommendations
machine learning and data mining, which is used in search of that handle the problem of scalability and sparsity posed by CF
similarities between items and customer preferences. successfully [8, 9, 10, 11, 12].
Recommendation techniques are categorized into three types:
collaborative filtering, content-based techniques, and hybrid The experimental results in the proposed approach showed that
techniques [1]. SVD-based CF achieved better recommendations compared to
item-based CF; the two achieved 1.0187 and 1.1969 in terms of
Collaborative filtering (CF) is the most successful technique in RMSE, respectively. They also achieved 0.8077 and 0.922 in
recommender systems. It recommends items by identifying terms of MAE, respectively. Our proposed approach also has
other users with similar tastes; uses their opinion to more accurate when compared to existing work with different
recommend items to the active user. CF systems have two main datasets.
approaches: memory-based and model-based approaches [1].
Memory-based approaches use user rating data to compute the This paper is organized as follows. Section 2 briefly describes
related works in the area of CF. The proposed approach is

9
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

described in Section 3. Sections 4 and 5 outline the proposed approach consists of two steps: memory-based and
experimental setting and the results, respectively. Conclusions model-based CF.
and suggestions for future work are presented in Section 6.

2.RELATED WORK
In this section, we will describe CF with two methods:
memory-based and model-based CF.
Jianfang and Pengfei in [13] introduced a CF algorithm
combined with the Singular Value Decomposition (SVD) and
Trust Factors (CFSVD-TF). For similarity computation, they
used the cosine distance metric. The dataset used was the
MovieLens 100k dataset containing 100,000 ratings (1-5) from
943 users for 1682 movies with each user having rated at least
20 movies. The proposed technique was evaluated using the
root mean square error (RMSE). The proposed method
obtained better prediction accuracy. It obtained 0.9762 in term
of RMSE with 10 neighbors.
Another method [14] applied CF based on items to produce a
recommendation in movies. The dataset was the Group Lens
M1, consisting of around one million ratings from 6,040 users
for 4,000 movies. For calculating the similarities between
movies, they used adjusted cosine similarity. The proposed
approach evaluated using MAE and achieved 0.938 in terms of
MAE with 20 neighbors.
The proposed approach in [15] introduced a book
recommendation system using item-based collaborative
filtering. Cosine distance metrics have been used to calculate
similarity books. The dataset used was goodbooks10k contains
ratings of 10,000 popular books and 53424 users. The proposed
method performed evaluations using MAE. The experimental
results achieved 0.72 in terms of MAE.
Fig 1: An Enhanced Collaborative Filtering-based
The proposed in [16] presented the Book Recommendation Approach for Recommender Systems
Algorithm using Deep Learning. The dataset used was
goodbooks10k contains 6 million ratings for 10,000 of the most 3.1 Memory-based Collaborative Filtering
popular books. The experiment randomly divided the data set Memory-based methods: are referred to as neighborhood-based
into an 80% training set and a 10% validation set and a 10% CF. The ratings of user-item combinations are predicted based
test set. The proposed technique was evaluated using RMSE. It on their neighborhoods. These neighborhoods can be defined in
obtained 1.1426 in terms of RMSE. one of two ways: User-based collaborative filtering, and Item-
based collaborative filtering. User-based collaborative filtering
Mala et al. [17] proposed a web-based movie recommender
calculates the similarity between users by comparing their
system that recommends movies to users based on their profile
ratings on the same item. Then computes the predicted rating
using the different recommendation algorithms such as K-
for an item. This method was initially quite popular. They are
Nearest Neighbor (KNN), singular value decomposition,
not easily scalable and sometimes inaccurate [2]. The
Alternating Least Squares (ALS) and Restricted Boltzmann
advantages of memory-based techniques are that they are
Machines (RBM). Experimental results showed that SVD
simple to implement. Other advantages are that the resulting
achieved better recommendations compared to KNN, ALS, and
recommendations are often easy to explain. On the other hand,
RBM. SVD, KNN, and ALS achieved 0.9002, 0.9375 and
memory-based algorithms do not work very well with sparse
1.069 in terms of RMSE respectively. They also achieved
rating matrices [2].
0.6925, 0.7263 0.9935 in terms of MAE respectively.
The K-nearest neighbors (KNN) Item-based CF finds the
Sandeep and Rajesh in [18] proposed a new method called
similarity between items by selecting the k most similar items.
Accelerated Singular Value Decomposition (ASVD). It uses
And their corresponding similarities are also determined using
momentum based Gradient Descent Optimization. They used
a cosine similarity measure. The prediction of the unknown
real world datasets (MovieLens100k, Film Trust and Yahoo
rating is created based on item-item similarity. The top items
Movie). The proposed technique evaluated using Absolute
are returned as recommendations. Figure 2 shows the
Error (MAE) and Root Mean Square Error (RMSE). The
Collaborative filtering Process.
experimental results showed that the proposed ASVD
outperformed other models of SVD using RSVD and SSVD. In the first step, the similarity between two items is measured
by computing the cosine of the angle between two vectors m
3.PROPOSED APPROACH ×n, m list of users and n list of items. The similarity between
The main stages for an enhanced collaborative filtering-based items i and j, given by [3]. Then the similar neighbor items are
approach for recommender systems are shown in Figure 1. The

10
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

found according to the following similarity. 𝑖.𝑗

sim(i,j)=𝑐𝑜𝑠(𝑖, 𝑗)=
𝑖 2∗ 𝑗 2

Fig 2: The Collaborative Filtering Process [3]

In The next step, the prediction on item i for a user u is non-negative elements.
obtained by computing the sum of the ratings given by the user
on the items similar to i. Each rating is weighted by the An m × m matrix U is called orthogonal if it equals to an m ×
corresponding similarity si,j between items i and j computed by m identity matrix. The diagonal elements in Σ (σ1, σ2, σ3, …
the following equation [3]: σn) are called the singular values of matrix A. Usually, the
singular values are placed in the descending order in Σ. The
𝑎𝑙𝑙 𝑠𝑖𝑚𝑖𝑙𝑎 𝑟 𝑖𝑡𝑒𝑚𝑠 ,𝑁(𝑠𝑖,𝑁∗𝑅𝑢 ,𝑁) column vectors of U and V are called the left singular vectors
Pu,i= 𝑎𝑙𝑙 𝑠𝑖𝑚𝑖𝑙𝑎𝑟 𝑖𝑡𝑒𝑚𝑠 ,𝑁( 𝑠𝑖,𝑁 ) and the right singular vectors respectively. SVD has many
Finally, the Top N items are selected using the computed desirable properties and is used in many important
similarly values. These items are not rated by the current user. applications. One of them is the low-rank approximation of
And recommended to the user. matrix A. The truncated SVD of rank k is defined [20, 21]:

Despite the success of the Item-Based CF technique, it has SVD (AK) =UK ΣKVKT
some problems such as sparsity and scalability [19]. To solve Where, UK ,VK are m × k and n × k matrices composed by the
these problems, we use model-based approach via Matrix first k columns of matrix U and the first k columns of matrix V
Factorization techniques as it deals with these problems respectively. K × k is the principle diagonal sub-matrix of Σ..
successfully and efficiently. AK represents the closest linear approximation of the original
matrix A with reduced rank k.
3.2 Model-based Collaborative Filtering
The model-based CF requires a learning phase in advance to 4. EXPERMINTAL WORK
learn a model to improve the performance of collaborative In the experimental work, we used the Large-scale Arabic
filtering. It includes some techniques such as clustering, Book Review (LABR) dataset. It has over 63K book reviews,
classification, latent model, Markov decision process (MDP), each with a rating of 1 to 5 stars [22].Table 1 is describing the
and matrix factorization. dataset used for testing the proposed approach. Figure 3 shows
Matrix Factorization is the most successful latent factor number of books for each rating.
models. It has become popular recently by combining good Table 1: Dataset used in proposed approach evaluation
scalability with predictive accuracy. It maps both users and
items to a joint latent factor space of dimensionality f. User- 63257
item interactions are modeled as inner products in that space Number of ratings
[20]. There are various matrix factorization models: Singular
Value Decomposition (SVD), Principal Component Analysis Number of unique book id's 2131
(PCA), Probabilistic Matrix Factorization (PMF) and Non-
Negative Matrix Factorization (NMF). 16486
Number of unique users
We use SVD as it is one of the most common and successful
matrix factorization techniques used in collaborative filtering.
60152
Singular Value Decomposition (SVD): is the powerful Number of unique reviews
technique of dimensionality reduction. This is a specific
implementation of the MF approach and is also related to the Average number of ratings per user 3.65
PCA. The main issue in SVD decomposition is to find a lower-
dimensional feature space [20]. SVD of an m × n matrix A is
of the form: Average number of ratings per book 28.23
SVD (A) =U ΣVT
Where, U and V are m × m and n × n orthogonal matrices
Average number of reviews per user 3.65
respectively. Σ is an m × n singular orthogonal matrix with

11
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

Cross-validation is a validation methodology for analyzing

statistical data. Cross-validation splits a dataset into k equally
Average number of reviews per book large partitions. One of the partitions is used as test partition
28.23
while the rest partitions are used as training partitions. The
algorithms then train a model with the training partitions and
when the training is complete, the model is tested with the test
partition, producing test data. This procedure continues until
every partition has been the test partition [23].
We divided the datasets into 80% for training and 20% for
testing data. Both KNN item-based CF and SVD based CF are
evaluated with 5-fold using the LABR dataset. The results are
analyzed, interpreted and compared using an absolute mean
error and a square root mean error. This will be seen in the
results section.

5. RESULTS
This section outlines the experiment results by presenting the
obtained MAE and RMSE values using the cross-validation
technique. Three experiments are carried out. In the first
experiment, we evaluated KNN Item-based CF. In the second
one, SVD-based CF is evaluated. In the third experiment,
performance comparisons with different methods are
performed.

Fig 3: Distribution of book ratings

5.1 KNN Item-based CF
In this experiment, the similarity between books is calculated
Only three fields were considered to predict user ratings using using a cosine similarity metric. We used the LABR dataset
collaborative filtering: user ID, book ID and rating. and cross-validated the algorithm with 10 as neighborhood
size. We run the experiment on the training data and used a test
To evaluate the overall performance, we used statistical
set to compute MAE and RMSE. Table 2 and Figure 4 show
accuracy metrics that are the most common evaluation measure
the average values in terms of RMSE and MAE are 1.1969,
for prediction accuracy
0.922 respectively.
Statistical accuracy metrics evaluate the accuracy of the
Table 2: Results of KNN Item-based CF
system by comparing the numerical recommendation scores
with the actual user ratings for the user-item pairs in the test
Fold Fold Fold Fold Fold Mea
dataset [3].
1 2 3 4 5 n
Mean Absolute Error (MAE) is a metric used to compute the
average of all the absolute value differences between the RM 1.19 1.19 1.19 1.20 1.19 1.19
prediction of the algorithm and the real rating [23]. The lower SE 9 0 6 6 3 7
the MAE the better the accuracy. In general, MAE can range
from 0 to Infinity, where Infinity is the maximum error
depending on the rating scale of the measured application MA 0.92 0.92 0.92 0.93 0.91 0.92
E 1 1 2 2 4 2
[3,24]. It is computed as follows:
1 𝑛
MAE=𝑛 𝑖=1 𝑝𝑖 − 𝑞𝑖

Where, KNN Item-based CF

pi is the actual rating 1.5
qi is the predicted rating
1
n is the amount of ratings
Root Mean Squared Error (RMSE): computes the mean value 0.5
of all the differences squared between the true and the
predicted ratings. Then, it proceeds to calculate the square root 0
out of the result. RMSE metric is the most valuable metric
when significantly large errors are unwanted [23,24]. It is Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Mean
computed as follows: RMSE MAE
1 𝑛 2
RMAE= 𝑛 𝑖=1( 𝑝𝑖 − 𝑞𝑖 ) Fig 4: Results of KNN Item-based CF

12
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

5.2 SVD-based CF 1.4

This experiment presented SVD based CF. The cross-validated
over the LABR dataset. We run the experiment on the training 1.2
data and used a test set to compute RMSE and MAE. Table 3,
1
Figure5 showed the RMSE and MAE scores. It achieved an
average of 1.0187, 0.8077 in terms of RMSE and MAE 0.8
respectively.
0.6
Table 3: Results of SVD-based CF
0.4
Fold Fold Fold Fold Fold Mea 0.2
1 2 3 4 5 n
0
RMS 1.02 1.00 1.02 1.01 1.02 1.01 RMSE MAE
E 1 8 0 9 3 9
KNN Item-based CF SVD-based CF
MAE 0.81 0.80 0.80 0.80 0.81 0.80
3 1 9 4 0 8
Fig 6: Performance of various CF

5.3 Performance Comparisons with

Different Methods
In this section, we present the RMSE and MAE values of each
method in the proposed approach and previous works with
SVD-based CF different datasets. To compare with existing work, we also
applied the proposed approach to the data set used in their
1.2 work.
1 Table 4: Performance Comparison Results
0.8 Method RMS MAE Datasets
0.6 E

0.4
0.976 -
Proposed in [13]
0.2 2
MovieLens
0 The proposed 0.936 - 100k
Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Mean approach 5

Proposed in [14] - 0.938

RMSE MAE The proposed - 0.730 MovieLens

approach 1M

Proposed in [15] - 0.72

Fig 5: Results of SVD-based CF
The proposed - 0.67 Goodbooks10k
We compared the best results achieved by the values of RMSE approach
and MAE. For the memory-based and model collaborative
filtering. Figure 6 shows the lowest RMSE and MAE values 1.142 -
obtained from SVD based CF comparing by KNN item-based Proposed in [16]
6
CF that indicates to superior the SVD based CF. Goodbooks10k
The proposed 0.843 -
approach 5

Proposed in [18] 3.432 3.228

The proposed 0.936 0.738 MovieLens
approach 5 100k

Proposed in [18] 3.154 2.955

The proposed 0.806 0.625 FilmTrust
approach

13
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

4 problem of scalability and sparsity in KNN Item-based CF. The

RMSE proposed approach when compared with different methods, it
3.5 gave the minimum RMSE, MAE values for the rating
3 predictions.
Our future work will focus on experimental evaluating of the
2.5
proposed collaborative filtering based on sentiment analysis.
2 We will, therefore, evaluate the accuracy and performance of
the proposed approach in Arabic datasets.
1.5
7. REFERENCES
1 [1] F. O. Izakaya, Y. O. Foliumin, and B. A. Ojokoh, 2015.
0.5 “REVIEW Recommendation systems: Principles, methods
and evaluation”. Egyptian Informatics Journal, 261–273.
0
[2] Charu C. Aggarwal, 2016, “Recommender Systems”, The
Proposed in Proposed in Proposed in Proposed in Textbook, ISBN 978-3-319-29659-3 (eBook), Springer
[13] [16] [18] [18] International Publishing Switzerland.
[3] Sarwar B, Karypis G, Konstan J. and Riedl J., 2001,
Existing approcch Our approach “Item-based collaborative filtering recommendation
algorithms Proceedings of the 10th international
conference on World Wide Web, 285-295.
Fig 6: Performance Comparison Results [4] Linden. G, Smith. B and York J., 2003, Amazon.com
Recommendations: Item-to-Item Collaborative Filtering”,
MAE IEEE Internet Computing, 76–80 .
3.5 [5] P. Prabhu and N. Anbazhagan, 2013, “FI-FCM Algorithm
for Business Intelligence”, Springer International
3 Publishing Switzerland. 518–528.

2.5 [6] Sarwar. B, Karypis.G, Konstan.J and Riedl.J. 2000

“Application of Dimensionality Reduction in
2 Recommender System - A Case Study”, in ACM
WEBKDD Workshop.
1.5
[7] Thi Do. M, Nguyen.D, and Nguyen.L., 2010, “Model-
1 based Approach for Collaborative Filtering”, The 6th
International Conference on Information Technology for
0.5 Education, 217-228.

0 [8] Vozalis. M, Angelos Markos A., and Margaritis K., 2014,

“Collaborative Filtering through SVD-Based and
Proposed in Proposed in Proposed in Proposed in Hierarchical Nonlinear PCA”, ICANN 2010, Springer-
[14] [15] [18] [18] Verlag Berlin Heidelberg, 395–400.
[9] Bokde. D. Sheetal Girase. Sh, Mukhopa. D., 2015,
Existing approcch Our approach
“Matrix Factorization Model in Collaborative Filtering
Fig 7: Performance Comparison Results Algorithms: A Survey”, Elsevier, the 4th International
Conference on Advances in Computing, Communication
As presented in Table 4, Figure6 and Figure7 the proposed and Control. Volume 49,136-146.
approach has better performance and accuracy as compared to
other methods on different datasets. [10] Hussein M., Okeyo G. and Mwangi.W, 2018, “Matrix
Factorization Techniques for Context-Aware
6. CONCIUSION Collaborative Filtering Recommender Systems: A
This paper proposed a system for predicting user preferences Survey”, Computer and Information Science; Vol. 11, No.
for items using two types of models: memory-based and 2; ISSN 1913-8989 E-ISSN 1913-8997.
model-based collaborative filtering. We used the LABR [11] Bhavanaa P, Kumarb V., and Padmanabhana V., 2019,
dataset. Also, we used four different datasets to compare the “Block based Singular Value Decomposition approach to
approach with existing work. The experimental results show matrix factorization for recommender systems” Pattern
that the proposed approach significantly improves the quality Recognition Letters journal homepage (Elsevier).
and accuracy of the recommendations in terms of RMSE and
MAE. The results also showed that SVD-based collaborative [12] Yuan X., Han L., Qian S., Guoxia Xu, and Yan H., 2019,
filtering is superior to the performance of KNN Item-based CF. “Singular value decomposition-based recommendation
The SVD-based CF approach was successful in addressing the using imputed data”, Knowledge-Based Systems Volume

14
 International Journal of Computer Applications (0975 – 8887)
Volume 176 – No. 41, July 2020

163, 485-494. University – Computer and Information Sciences,1-5.

[13] Wang J., HanP., MiaoY. and Zhang F., 2019, “A [19] Mohamed M., Khafagy M., Ibrahim M, 2019,
Collaborative Filtering Algorithm Based on SVD and “Recommender Systems Challenges and Solutions
Trust Factor”, International Conference on Computer, Survey”, International Conference on Innovative Trends
Network, Communication and Information Systems .33- in Computer Engineering (ITCE’2019),149-155
39.
[20] Bokde D., Girase S., and Mukhopadhyay D., 2014, “Role
[14] Ponnam L. and Punyasamudram S., 2016, “Movie of Matrix Factorization Model in Collaborative Filtering
Recommender System Using Item Based Collaborative” Algorithm: A Survey”, International Journal of Advance
Filtering Technique”, Proceedings of ICETETS 2016, Foundation and Research in Computer (IJAFRC) Vol., 1.
Kings College of Engineering, 56-60.
[21] Polat H. and Du W. 2005, “SVD-based Collaborative
[15] Kaivan Shah, 2019, “Book Recommendation System Filtering with Privacy”, ACM Symposium on Applied
using Item based Collaborative Filtering”, International Computing SAC’05, Santa Fe, New Mexico, USA, ACM
Research Journal of Engineering and Technology 1581139640/05/0003, ©ACM, 13-17.
(IRJET), 5960-5965.
[22] Nabil M., Aly M., Atiya A., 2013, “LABR: A Large-
[16] Liu A. S, Gao J., 2019, “Book Recommendation Scale Arabic Book Reviews Dataset”, Aclweb. Org. 494–
Algorithm Based on Deep Learning”, International 498 (2013).
Journal of Science, 152-156.
[23] Najafi S., Salam Z., 2016, “Evaluating Prediction
[17] Saraswat M., Anil Dubey A., Naidu S., Vashisht R. and Accuracy for Collaborative Filtering Algorithms in
Singh A., 2020, “Web-Based Movie Recommender Recommender Systems”, KTH Royal Institute of
System”, Springer, Ambient Communications and Technology School of Computer Science and
Computer Systems, 291-301. Communication.
[18] Raghuwanshi S., Pateriya R., 2018,” Accelerated Singular [24] Herlocker J., Konstan J., Terveen L., and Riedl J., 2004,
Value Decomposition (ASVD) using momentum based “Evaluating collaborative filtering recommender
Gradient Descent Optimization”, Journal of King Saud systems”, ACM Trans. Inf., 5–53.

IJCATM : www.ijcaonline.org 15

View publication stats