The CSI Journal on

Computer Science and Engineering

Vol. 17, No. 2, 2020
Pages 17-25
Regular Paper

Hybrid Movie Recommender System based on Resource

Allocation
Mostafa Khalaji1 Chitra Dadkhah1 Joobin Gharibshah2

1
Faculty of Computer Engineering, K. N. Toosi University of Technology, Tehran, Iran
2
University of California – Riverside, CA

Abstract
Recommender Systems are inevitable to personalize user’s experiences on the Internet. They are using different approaches to recommend the
Top-K items to users according to their preferences. Nowadays recommender systems have become one of the most important parts of large-
scale data mining techniques. In this paper, we propose a Hybrid Movie Recommender System (HMRS) based on Resource Allocation to
improve the accuracy of recommendation and solve the cold start problem for a new movie. HMRS-RA uses a self-organizing mapping neural
network to clustering the users into N clusters. The users' preferences are different according to their age and gender, therefore HMRS-RA is
a combination of a Content-Based Method for solving the cold start problem for a new movie and a Collaborative Filtering model besides the
demographic information of users. The experimental results based on the MovieLens dataset show that the HMRS-RA increases the accuracy
of recommendation compared to the state-of-art and similar works.
Keywords: hybrid recommender system, self-organizing map, content-based method, resource allocation, collaborative filtering

1. Introduction the other users in the system. 2: Content-Based Method

(CBM) in which the recommendation is based on active user
Facing the huge amount of internet information and history and the descriptive features of items. 3: Knowledge-
increasing online shopping, the need for recommendation Based Systems (KBS), in which the recommendation is based
systems became obvious as a way to guide users toward their on explicit knowledge regarding users’ requirements. Finally,
preferences. In the recommender systems, users are the set of Hybrid Systems (HS) are a combination of different
the actors in the system who have interactions with each other approaches to improve the performance of the system [1,4].
and items are the set of entities that can be liked or observed Meta heuristic-based recommender systems aim to find a
by users [1]. In recommender systems, when the number of possible solution for complex optimization tasks. In these
users or items grows exponentially at an enormous speed, systems, a set of possible solutions search through the solution
scalability becomes a problem, so a recommendation method space to find the optimal region of the search space and
should be fast and efficient on a small dataset as well as a large discover a near-optimal solution in a reasonable time. For
one. On the other hand, lacking the information of new users achieving this aim, a set of search operators are designed in
or items causes the cold-start problem which affects the such a way that the most successful solutions are let in to be
accuracy of prediction. One of the techniques to handle this evolved through the next iterations.
problem was introduced by Liben-Nowell and Kleinberg in Resource Allocation (RA) is a measure used to compute
2007 [2,3] which find a similarity between two entities in a the closeness of nodes based on their shared neighbors. The
network. Resource Allocation algorithm was introduced in 2009 by Tao
There are three foundation approaches in recommender Zhou et al as part of a study to predict links in various
systems [1]: 1: Collaborating Filtering (CF) which the networks [5].
recommendation is based on the performance or behavior of
M. Khalaji, C.Dadkhah & J.Gharibshah: Hybrid movie recommender system based on resource allocation (Regular Paper) 18

According to the RA algorithm, a pair of nodes––which are aspects: (1) emotional offset of user reviews can be unveiled
not connected directly––can share their resources through by sentiment analysis and can be utilized to revise original
their common neighbors. RA-reliability indicates the degree user ratings; (2) user preference and doctor features are
of reliability for the estimated similarity between the users 𝑢𝑖 extracted by Latent Dirichlet Allocation and incorporated into
and 𝑢𝑗 [6]. RA is used to find out missing relations or possible conventional matrix factorization [14]. Jamali et al. proposed
future relations in the network [4], e.g., recommend friends on a model-based approach for the recommendation in social
social networks [7] or recommend an item in online shopping networks, employing matrix factorization techniques and
[8]. incorporated the mechanism of trust propagation into the
The neural network has the advantages of strong feature model. Trust propagation is a crucial phenomenon in the social
extraction, effective learning, and simple process. Self- sciences, social network analysis and trust-based
Organizing Map (SOM) is an unsupervised neural network recommendation [15]. Li et al. proposed a novel
and it is more suitable for the development of complex and recommendation method called TruCom. In a multi-category
different recommender systems. We could cluster similar item recommendation domain, TruCom, first generates a
users or items in the same feature space area using SOM domain-specific trust network pertaining to each domain and
neural network [9]. then builds a unified objective function for improving
In this paper, we propose a hybrid Movie Recommender recommendation accuracy by incorporating the hybrid
System based on Resource Allocation (HMRS-RA) to handle information of direct and indirect trust into a matrix
scalability, cold-start problems, and increases the accuracy of factorization recommendation model [16]. Guo et al. proposed
recommendation. TrustSVD, which is a trust-based matrix factorization
The novelty of our approach is the use of RA for technique. By analyzing the social trust data from four real-
computing the weights for user similarity in person similarity world data sets, they have concluded that not only the explicit
measure to increase the precision of the predictions. rating but also the implicit influence of trust should be
The structure of this paper is organized as follows: Section 2 considered in a recommendation model [17]. Al Hasan et al.
presents the summary of various recent research. Section 3 gathered some representative link prediction methods
describes the structure of HMRS-RA. In Section 4, we according to the type of the models. They have considered
demonstrate the results of the implementation and evaluation three types of models: 1: the traditional (non-Bayesian) model
of HMRS-RA. Finally, the conclusion is presented in Section which extracts a set of features to train a binary classification
5. model. 2: the probabilistic model which models the joint-
probability among the entities in a network using Bayesian
graphical models. 3: the linear algebraic model computes the
2. Related Work similarity between the nodes in a network by rank-reduced
similarity matrices [18]. Xia et al. proposed a recommendation
As mentioned in the abstract section, HMRS-RA combines
algorithm Improved Weighted Network-Based Inference
the CF and CBM to improve the efficiency of recommender
(INBIw) that improves the original weighted network-based
systems, so we have studied the researches in three categories:
inference by introducing a tunable parameter 𝛽 to reduce the
collaborative filtering RS, content-based RS, and hybrid RS.
influence of high-degree nodes. In order to evaluate the
Because recent researches in CFRS focus on Metaheuristic,
recommendation performance of INBIw, ranking position rate
we have also explained these researches. We have
and hitting rate are calculated [19]. Zhao et al. proposed a
summarized the researches in Fig.1.
method for supporting resource allocation in business process
management. In their method, resource allocation is
2.1. CF approach considered as a multi-criteria decision problem and solved by
We divided the researches in CFRS into probabilistic and a new entropy-based clustering ensemble approach. By
trust approaches. In the probabilistic approach, Ma et al. mining resource characteristics and task preference patterns
proposed a factor analysis approach that is called Sorec, which from past process executions, the right resources could be
is based on probabilistic matrix factorization to solve the data recommended to improve resource utility [20]. Lu et al.
sparsity by employing both users' social network information summarize link prediction algorithms, emphasizing the
and rating records [10]. Chaney et al. developed Social contributions from physical perspectives and approaches, such
Poisson Factorization (SPF), a probabilistic model that as the random-walk-based methods and the maximum
incorporates social network information into a traditional likelihood methods. They also introduced three typical
factorization method; SPF introduces the social aspect to applications: reconstruction of networks, evaluation of
algorithmic recommendation [11]. Chen et al. proposed a network evolving mechanism, and classification of partially
probabilistic recommender system that uses the clustering labeled networks [21]. Fangyi Hu introduced a three-segment
method. Their system learns co-preference patterns from similarity measure method for the collaborative filtering
historical transaction data and recommends items accordingly model. He improved the performance of similarity measure by
[12]. computing the similarity between users based on the number
In the trust approach, we divided the researches into matrix of user ratings along with item similarity and user attribute
factorization and link prediction methods. In the matrix similarity [22].
factorization method, Ma et al. proposed a novel probabilistic
factor analysis framework, which naturally fuses the users'
2.1.1. Clustering approach
tastes and their trusted friends' favors together [13]. Zhang et
In this section, we have explained the methods used in the
al. developed a novel healthcare recommendation system
papers that focus on clustering in model-based CF. Belacel et
called iDoctor, which is based on hybrid matrix factorization
al. introduced a scalable recommender system based on a
methods. iDoctor differs from previous work in the following
collaborative filtering approach. They improved the time and
The CSI Journal on Computer Science and Engineering, Vol. 17, No. 2, 2020 19

accuracy of their proposed system using the split-merge improving the accuracy of prediction of unobserved movies
clustering algorithm [23]. Kant et al. introduced a method to for active users. NWS_RS managed the scalability problem
determine the selection of the center of the cluster in the K- and solved the data sparsity problem [34]. Khalaji et al
means clustering operation. Their method was able to solve proposed a new recommender system called CUPCF which
the data sparsity problem [24]. Wang et al. introduced a new was a combination of two similarity measures in CF to solve
method called the CDIE. They used the Co-Clustering method the data sparsity and better recommendation. CUPCF used
to learn cross-domain comprehensive representations of items two similarity measures simultaneously as a new method for
by collectively leveraging single-domain and cross-domain decreasing the error rate of the system [35].
sessions within a unified framework. Their method solved the
data sparsity problem [25]. Rafiee et al. proposed a similarity- 2.2. Content-based approach
based link prediction algorithm, referred to as CNDP, which In content-based methods, Mooney et al. proposed a
in this algorithm the similarity score is determined according content-based book recommender system for text
to the structure and specific characteristics of the network, as categorization which their approach has the advantage of
well as the topological characteristics. In their proposed being able to recommend previously unrated items to users
method, a new metric for link prediction is introduced, with unique interests and to provide explanations for its
considering clustering coefficient as a structural property of recommendations [36]. Deldjoo et al. proposed a new content-
the network. Moreover, their method also considers the based recommender system that encompasses a technique to
neighbors of shared neighbors of each pair of nodes, which automatically analyze video contents and to extract a set of
leads to achieving better performance than the other similar representative stylistic features (lighting, color, and motion)
link prediction methods [26]. Zhu et al. proposed link grounded on existing approaches of Applied Media Theory
prediction indices based on both Network Structure and Topic [37]. Van den Oord et al. use a latent factor model for the
Distribution (NSTD). Their approach makes full use of the recommendation, and predict the latent factors from music
network characteristics, such as homophily, transitivity, audio when they cannot be obtained from usage data [38].
clustering, and degree heterogeneity. And they combined Yang et al. proposed a movie recommendation system
these characteristics with topic similarity when constructing according to scores that the users have provided. In view of
indices based on both directly and indirectly connected nodes the movie evaluation system, the impacts of access control and
[27]. Liu et al. introduced a novel Collaborative Linear multimedia security are analyzed, and secure hybrid cloud
Manifold Learning (CLML) algorithm which can optimize the storage architecture is presented. Mobile-Edge Computing
consistency of nodes similarities using the manifolds (MEC) technology is used in the public cloud which
embedded between the target and the auxiliary network [28]. guarantees the high-efficiency requirements of the
Mazzouzi et al. proposed a new effective recommender transmission of the multimedia content. The processes of the
system for TED (Technology, Entertainment, and Design) system include registration, user login, role assignment, data
talks that first groups the users according to their preferences encryption, and data decryption [39]. Wang et al. proposed a
and then provides a powerful mechanism to improve the content-based recommender system for computer science
quality of recommendations for users. In their system, the publications. Their system recommends suitable journals or
authors used the Pearson Correlation Coefficient (PCC) conferences with a priority order based on the abstract of a
method and TED talks to create the TED user-user matrix. manuscript. To follow the fast development of computer
Then, they used the k-means clustering method to group the science and technology, a web crawler is employed to
same users in clusters and create a predictive model. Finally, continuously update the training set and the learning model.
they used this model to make relevant recommendations to To achieve the interactive online response, they propose an
other users [29]. Xiaopan et al. for solving the data sparsity efficient hybrid model based on chi-square feature
problem in CF, proposed a SOM clustering collaborative selection and softmax regression [40]. Rahimpour et al.
filtering algorithm based on Singular Value Decomposition introduced a new method for a content-based filtering
(SVD) which reduces the dimensions of the original matrix. recommender system. They use the interactions of each user
by decomposing to the item and user latent factor [30]. and analyze them to propose a new user model and capture
Parvin et al. proposed a novel CF method for predicting user’s interests. Their system built the user model based on a
missing ratings accurately. Their proposed method, called Bayesian framework called the Dirichlet Process Mixture
TCFACO (Trust-aware Collaborative Filtering Ant Colony Model. They improved the accuracy of their system in
Optimization), used trust statements as a rich side information comparison to other methods [41].
with Ant Colony Optimization (ACO) method [31]. For
increasing the accuracy of the recommendation of user-based, 2.3. Hybrid approach
Tohidi et al. proposed a hybrid approach based on clustering In hybrid systems, Lee et al. proposed a new recommender
and evolutionary algorithm. They combined the K-means system that combines collaborative filtering with Self-
clustering method along with two Metaheuristic algorithms Organizing Map (SOM) neural network. First, all users are
such as FOA and APSO [32]. Khodaverdi et al. proposed a segmented by demographic characteristics and users in each
movie hybrid recommender system based on clustering and segment are clustered according to the preference of items
popularity. Their system clusters the users who were similar using the SOM neural network [42]. Nadi et al. proposed a
to each other by using the K-means clustering method and uses fuzzy recommender system (FARS) based on the
rating popularity to predict the users' preferences for specific collaborative behavior of ants. FARS works in two phases:
movies [33]. Khalaji proposed a new recommender system modeling and recommendation. First, user's behaviors are
called NWS_RS for movie recommendations. His method was modeled offline and the results are used in the second phase
able to personalize the recommendation by segmenting users' for the online recommendation. Fuzzy techniques provide the
age. NWS_RS used the New Weighted Similarity (NWS) for
M. Khalaji, C.Dadkhah & J.Gharibshah: Hybrid movie recommender system based on resource allocation (Regular Paper) 20

possibility of capturing uncertainty among user interests and

Ant-based algorithms optimize the solutions for predicting
phase. The performance of FARS is evaluated using log files
of Information and Communication Technology Center of
Isfahan municipality in Iran and have compared with Ant-
based Recommender System (ARS) [43]. Roh et al. proposed
a three-step CF recommendation model, which is composed
of profiling, inferring, and predicting steps while considering
prediction accuracy and computing speed simultaneously.
Their model combines a CF algorithm with two machine
learning methods, Self-Organizing Map (SOM) and Case-
Based Reasoning (CBR) by changing an unsupervised
clustering problem into a supervised user preference reasoning
problem, which is a novel approach for the CF Fig. 1: The classification of related work
recommendation field [44]. May et al. proposed a neural
networks-based clustering collaborative filtering algorithm in 3. Proposed HMRS-RA
the e-commerce recommendation system. Their algorithm
tries to establish a classifier model based on Back Propagation HMRS-RA consists of two phases: online and offline. In
(BP) neural network for the pre-classification of items. They the offline phase, the preprocessing of the data is done and
analyzed and discussed their algorithm from multiple aspects then the recommendation is made on the online phase as is
[45]. Kim et al. proposed a robust document context-aware depicted in Fig. 2.
hybrid method, which integrates Convolutional Neural In the offline phase, HMRS-RA filters users based on their
Network (CNN) into Probabilistic Matrix Factorization gender (feminine and masculine) and age (range of 20-39
(PMF). Their method captured contextual information using years and range of 40-60 years). In this step, four rating
the statistics of items [46]. Katarya et al. proposed a matrices ( 𝑅𝑚𝑎𝑙𝑒 , 𝑅𝑓𝑒𝑚𝑎𝑙𝑒 , 𝑅𝑎𝑔𝑒(20−39) , 𝑅𝑎𝑔𝑒(40−60) ) are
component of Hybrid Music Recommender Systems (HMRS), generated according to the original rating matrix, which
which combined context-sensitive and collaborative filtering consists of the rating of each user for observed movies. Where,
approaches. Their method used the timestamp of user rating 𝑅𝑚𝑎𝑙𝑒 indicates groups of male and 𝑅𝑎𝑔𝑒(40−60) represents the
for modeling user behaviors. They used the Depth-First- rating of the users who are older than 40 years old and below
Search (DFS) algorithm which traverses the whole graph 60. The ratings are from the set {1, 2, 3, 4, 5} in which a rating
through the paths in different contexts and generated the of 1 indicates an extreme dislike and a rating of 5 indicates the
ranked list of recommended items using the Bellman-Ford extreme like for a movie.
algorithm with multi-layer context graph [47]. De Pessmier et HMRS-RA identifies sets of similar users based on movies
al. proposed a recommender system that offers personalized rating using the Self-Organizing Map (SOM) clustering
recommendations for travel destinations to individuals and method and uses N clusters to build up a model to predict the
groups. These recommendations are based on the users' rating rating of unobserved movies for active users in the online
profile, personal interest, and specific demands for their next phase. For each category of users, HMRS-RA defines the most
destination. Their recommendation algorithm was a hybrid preferred genres of movies rated by users. We consider the
approach that combined content-based collaborative filtering five popular genres of the movie, such as Action, Adventure,
and knowledge-based models. For groups of users, such as Comedy, Drama and Romance.
families or friends, individual recommendations are suggested At the end of the offline phase, four rating
′ ′ ′ ′
into group recommendations, with an additional opportunity matrices: 𝑅𝑚𝑎𝑙𝑒 𝑐1
, 𝑅𝑚𝑎𝑙𝑒 𝑐2
, 𝑅𝑓𝑒𝑚𝑎𝑙𝑒𝑐1
, 𝑅𝑓𝑒𝑚𝑎𝑙𝑒 𝑐2
with 𝑚 × 5
for users to give feedback on these group recommendations. dimensions where m is the number of users and 5 is the
A group of test users has evaluated the recommender system ′
number of genres will generate. Each entry of 𝑅𝑚𝑎𝑙𝑒 𝑐1
and
using a prototype web application [48]. Wei et al. proposed a ′
𝑅𝑚𝑎𝑙𝑒𝑐2 is the average of the rating of the observed movies
hybrid movie recommendation approach using tags and
that belong to five popular genres for masculine users in
ratings. First, they constructed social movie networks and a
Clusters.
preference-topic model. Then, they extracted, normalized, and
reconditioned the social tags according to user preference
based on social content annotation. Finally, they enhanced the
recommendation model by using supplementary information
based on user historical ratings [49]. Deldjoo et al. proposed
multimedia recommender systems called MMRS. They
combined content-based and collaborative filtering
approaches. The target of their system was the
recommendation of music, movies and images using deep
learning and feature extraction [50]. Tarus et al. proposed a
new hybrid recommender system for e-learning. Their system
used sequential pattern mining called SPM along with context-
awareness and collaborative filtering approach for suggesting
learning resources to the users. They improved the quality and
accuracy of their system [51].
Fig. 2: The HMRS-RA structure
The CSI Journal on Computer Science and Engineering, Vol. 17, No. 2, 2020 21

On the other side, another four rating 4. Evaluation of HMSRS-RA

′ ′ ′
matrices: 𝑅𝑎𝑔𝑒(20−39 )𝑐1
, 𝑅𝑎𝑔𝑒(20−39)𝑐2
, 𝑅 𝑎𝑔𝑒(40−60) 𝑐1
,
′
𝑅𝑎𝑔𝑒(40−60)𝑐2 , with 𝑚 × 5 dimensions where m is the number The effectiveness of HMRS-RA is evaluated on
of users and 5 is the number of genres will generate. Each MovieLens [52] data set which consists of 943 users, 1682
′ ′ items with 100,000 user ratings for movies. The ratings are
entry of 𝑅𝑎𝑔𝑒(20−39 )𝑐1
and 𝑅𝑎𝑔𝑒(20−39) 𝑐2
are the average of the
from the set {1, 2, 3, 4, 5} that indicates the level of like or
rating of the observed movies that belong to five popular
dislike of the observed movies.
genres for users who are older than 19 years and below 40 in
For evaluating the HMRS-RA, we use the five-fold cross-
Clusters.
validation algorithm. The cross-validation procedure consists
In the online phase, first HMRS-RA finds the cluster that
of 5 iterations and in each iteration, 80 % of the data and the
the active user belongs to using a three layers Artificial Neural
rest of the data (20%) is considered as training and test data,
Network (ANN) classification and then detects preferred
respectively [53]. The initial weight values in the SOM
genre using CBM (Content-Based Method) and predicts the
clustering method consider randomly, so each fold is repeated
rating of unobserved movies for the active user using CF
10 times with independent running. We calculated the average
method. After determining the active user cluster, the
value of the Mean Absolute Error (MAE) using Eq. (4) over
similarity between the active user and the users in his/her
different iterations of cross-validation.
cluster is calculated by Pearson similarity measure as shown
in Eq. (1) to determine the k-neighbors of the active user.
∑𝑛𝑖=1 |𝑟̂𝑢,𝑖 − 𝑟𝑢,𝑖 |
𝑀𝐴𝐸 = (4)
𝑆𝑖𝑚(𝑢, 𝑣) 𝑛
∑𝑘∈𝐼𝑢 ∩𝐼𝑣(𝑅𝑎𝑡𝑒𝑢𝑘 − µ𝑢 ) . (𝑅𝑎𝑡𝑒𝑣𝑘 − µ𝑣 ) Where, 𝑟̂𝑢,𝑖 is the predicted rating value of movie 𝑖 by user
= (1) 𝑢 with HMRS-RA, 𝑟𝑢,𝑖 is the actual rating value given by user
√∑𝑘∈𝐼𝑢∩𝐼𝑣 (𝑅𝑎𝑡𝑒𝑢𝑘 − µ𝑢 )2 . √∑𝑘∈𝐼𝑢 ∩𝐼𝑣(𝑅𝑎𝑡𝑒𝑣𝑘 − µ𝑣 )2
𝑢 for movie 𝑖 and 𝑛 is the number of the predicted values.
Where K is the set of the same movies observed by user u In our experiment, first, we estimated the real rating value
for each genre by calculating the average rating value of
and v. 𝑅𝑎𝑡𝑒𝑢𝑘 indicates the rating of active user 𝑢 for movie 𝑘
observed movies of each genre as shown in Table 1, for the
and µ𝑢 is the average movie rating given by user u. The value
active user. Then we predicted the rating of the genre that its
of 𝑆𝑖𝑚(𝑢, 𝑣) always lies in the range [-1, 1]. A value 1
real rating is more than value 4 by HMRS-RA and calculated
indicates the most similarity between users while a value -1
the MAE as shown in Table 2. We considered the genre which
indicates users are not similar. The similarity between two
has the highest and lowest MAE value as the worst case and
users who have rated the same unpopular movies is stronger
the best case, respectively. We predicted the rating of movies
than the similarity between users who have rated the same
of the worst and best case by HMRS-RA and calculate the
popular movies. Therefore HMRS-RA calculates the weight
MAE as shown in Table 3 based on the initial rating matrix.
for the similarity between the users 𝑢𝑖 and 𝑢𝑗 using the
Finally, the overall MAEs for the worst and the best cases are
Resource Allocation (RA) weighting method as shown in Eq. calculated by multiplication of MAEs in Table 2 and Table 3
(2). [54] as shown in Table 4.

1 Table 1: The genre ratings by an active user

𝑅𝑅𝐴 (𝑢𝑖 , 𝑢𝑗 ) = ∑ (2)
𝑘𝑧 Genre Adventure Romance Comedy Drama Action
𝑧∈(𝑢𝑖 )∩(𝑢𝑗 ) Rate 4 4 3 1 4

Where 𝑘𝑧 is the number of users that rated the movie z. Table 2: The MAE value for preferred genres for an active user
(𝑢𝑖 ) is the set of neighbors of user 𝑢𝑖 that rated movie z. So Genre Adventure Romance Action
MAE 0.399 0.520 0.481
HMRS-RA predicts the rating of the unobserved movie 𝑖 for
active user 𝑢 using Eq. (3).
Table 3: The MAE value of the worst and the best case for an
active user
𝑃𝑟𝑒𝑑𝑖𝑐𝑡(𝑢, 𝑖) Genre Worst case Best case
= µ𝑢
MAE 0.71 0.60
∑𝑚𝑗=1(𝑟𝑣𝑗 ,𝑖 − µ𝑣 ). 𝑆𝑖𝑚(𝑢, 𝑣𝑗 ). 𝑅𝑅𝐴 (𝑢, 𝑣𝑗 )
+ (3)
∑𝑚𝑗=1 |𝑆𝑖𝑚(𝑢, 𝑣𝑗 ). 𝑅𝑅𝐴 (𝑢, 𝑣𝑗 )|
Table 4: Overall MAE of HMRS-RA for an active user
Genre Worst Case Best Case

Where 𝑚 is the number of neighbors for active user u and MAE 0.3692 0.2394
µ𝑢 is an average of observed movies rating of user 𝑢 and 𝑟𝑣𝑗 ,𝑖
indicates the rating done by user 𝑣𝑗 for movie 𝑖. We consider Table 5: The MAE of HMRS-RA after 50 iteration
half of the users in the active user's cluster for the value of m. MAE
HMRS-
After predicting the rating of all unobserved movies, HMRS- 20-39 40-60
RA Men Women
years years
RA recommends the top K-movies to active user u by ranking
Best case 0.16043 0.23653 0.16881 0.23584
the predicted rating values. Worst
0.30656 0.45268 0.29701 0.44450
case
overall 0.233495 0.344605 0.23291 0.34017
M. Khalaji, C.Dadkhah & J.Gharibshah: Hybrid movie recommender system based on resource allocation (Regular Paper) 22

Fig. 3 and Fig. 4 indicate the comparison of HMRS-RA Table 6: The evaluation of algorithms according to user
with the CF-RA in [6] and the RS that combined the traditional gender
CF and SOM methods according to the MAE criteria. The MAE
Method
results show the efficiency of our proposed algorithm and the Men Women
increasing accuracy of recommendation by HMRS-RA. K-Means Leader
0.74 0.74
[24]
Table 6 and Table 7 indicate the comparison of HMRS- K-Means [24] 0.755 0.755
RA with the latest work in [22], [24] and all of the methods in Three-Segment
recommender systems that have been cited in them on 0.75 0.75
Similarity [22]
recommender systems according to the MAE criteria. This BCF [22] 0.78 0.78
section focuses on various measures that are related to new NHSM [22] 0.83 0.83
user cold-start problems and they are Proximity-Impact- PIP [22] 0.86 0.86
Popularity (PIP), NHSM (New heuristic similarity measure), Cosine [22] 0.865 0.865
and other methods for computing users similarity such as Pearson [22] 0.87 0.87
Pearson, Cosine. BCF is a similarity measure based on HMRS-RA 0.233495 0.344605
Bhattacharyya coefficient [55]. The three-segment similarity
Table 7: The evaluation of algorithms according to user age
measure is a model for solving cold-start and data sparsity MAE
problems in the recommender system. K-means is a clustering Method
Ages 20 to 39 Ages 40 to 60
method in recommender systems and K-Means Leader is a K-Means Leader
new clustering collaborative framework, which improves the 0.74 0.74
[24]
quality of clustering and recommendations. K-Means [24] 0.755 0.755
Three-Segment
0.75 0.75
Similarity [22]
BCF [22] 0.78 0.78
NHSM [22] 0.83 0.83
0.8 PIP [22] 0.86 0.86
0.7 Cosine [22] 0.865 0.865
Pearson [22] 0.87 0.87
0.6
HMRS-RA 0.23291 0.34017
0.5
MAE

0.4
0.3
0.2
5. Conclusion
0.1 In this paper, we proposed a Hybrid Movie Recommender
0 System which combines collaborative filtering and content-
Men Women based filtering to solve the cold-start problem for new items.
By considering the contextual information such as genre,
HMRS-RA CF+SOM CF-RA HMRS-RA would diminish the cold start problem for new
movies according to their genre. The proposed method
(HMRS-RA) solves the scalability problem using clustering to
Fig. 3: The comparison of methods according to user gender
reduce the dimensionality of the data. By considering the
resource allocation as a weight for detecting the similarity
between users in each cluster, we improved the performance
of the recommendation comparing with a number of state-of-
the-art and latest work in recommender systems. The
0.8
experimental results showed that the MAEs of our proposed
0.7 algorithm are 0.23, 0.34, 0.23, and 0.34 for men, women, age
0.6 of 20-39 and age of 40-60, respectively. So, HMRS-RA
0.5 increased the accuracy of recommendation. In the future, we
MAE

0.4
would like to classify the users based on deep learning
approaches such as convolutional neural networks in the case
0.3 of a large dataset.
0.2
0.1 References
0
Ages of 20 to 39 Ages of 40 to 60
[1] C. Aggarwal, Recommender Systems. Springer
HMRS-RA CF+SOM CF-RA International Publishing, 2016.
[2] D. Liben-Nowell, and J. Kleinberg, "The link-prediction
problem for social networks", Journal of the American Society
Fig. 4: The comparison of methods according to user age for Information Science and Technology, Vol. 58, No. 7, pp.
1019-1031, 2007.
[3] Z. Wu, and Y. Li, "Link Prediction Based on Multi-steps
Resource Allocation", in Proceedings of the 2014
The CSI Journal on Computer Science and Engineering, Vol. 17, No. 2, 2020 23

IEEE/WIC/ACM International Joint Conferences on Web the Twenty-Ninth AAAI Conference on Artificial
Intelligence (WI) and Intelligent Agent Technologies (IAT), Intelligence, AAAI Press: Austin, Texas, pp. 123-129, 2015.
IEEE Computer Society, Vol. 01, pp. 355-360, 2014. [18] M.A. Hasan, and M.J. Zaki, "A Survey of Link Prediction
[4] G. Adomavicius, and A. Tuzhilin, "Toward the next in Social Networks," in Social Network Data Analytics, C.C.
generation of recommender systems: A survey of the state-of- Aggarwal, Editor, Springer US: Boston, MA, pp. 243-275,
the-art and possible extensions," IEEE transactions on 2011.
knowledge and data engineering, 17(6), pp.734-749, 2005. [19] J. Xia, F. Wu, Z. Xiong, M. Qiu, and C. Xie, "Modeling
[5] T. Zhou, L. Lü, and Y. Zhang, "Predicting missing links recommender systems via weighted bipartite network,"
via local information," The European Physical Journal B, Concurrency and Computation: Practice and Experience, Vol.
Vol. 71, No. 4, pp. 623-630, 2009. 29, No.14 pp. e3895-n/a, 2017.
[6] A. Javari, J. Gharibshah, and M. Jalili, "Recommender [20] W. Zhao, H. Liu, W. Dai, and J. Ma, "An entropy-based
systems based on collaborative filtering and resource clustering ensemble method to support resource allocation in
allocation," Social Network Analysis and Mining, Vol. 4, No. business process management," Knowledge and Information
1, pp. 234, 2014. Systems, Vol. 48, No.2, pp. 305-330, 2016.
[7] M. Roth, A. Ben-David, D. Deutscher, G. Flysher, I. Horn, [21] L. Lü, and T. Zhou, "Link prediction in complex
A. Leichtberg, and R. Merom, "Suggesting friends using the networks: A survey," Physica A: Statistical Mechanics and its
implicit social graph," in Proceedings of the 16th ACM Applications, Vol. 390, No. 6, pp. 1150-1170, 2011.
SIGKDD international conference on Knowledge discovery [22] F. Hu, "Three-Segment Similarity Measure Model for
and data mining, ACM: Washington, DC, USA., pp. 233-242, Collaborative Filtering," In International Conference on Data
2010. Mining and Big Data, Springer, Cham, pp. 138-148, 2018.
[8] M. Li, BM. Dias, I. Jarman, W. El-Deredy, and P.J. Lisboa, [23] N. Belacel, G. Durand, S. Leger, and C. Bouchard,
"Grocery shopping recommendations based on basket- "Scalable Collaborative Filtering Based on Splitting-Merging
sensitive random walk," in Proceedings of the 15th ACM Clustering Algorithm," In International Conference on Agents
SIGKDD international conference on Knowledge discovery and Artificial Intelligence, pp. 290-311, 2018.
and data mining, ACM: Paris, France, pp. 1215-1224, 2009. [24] S. Kant, T. Mahara, V.K. Jain, D.K. Jain, and A.K
[9] H. Ma, and H. Liu, "Design of Clothing Clustering Sangaiah, "LeaderRank based k-means clustering
Recommendation System on SOM Neural Network," In 8th initialization method for collaborative filtering," Computers
International Conference on Social Network, Communication & Electrical Engineering, Vol. 69, pp. 598-609, 2018.
and Education, Atlantis Press, 2018. [25] Y. Wang, C. Feng, C. Guo, Y. Chu, and J.N. Hwang,
[10] H. Ma, H. Yang, M.R. Lyu, and I. King, "SoRec: social "Solving the Sparsity Problem in Recommendations via
recommendation using probabilistic matrix factorization," in Cross-Domain Item Embedding Based on Co-Clustering,"
Proceedings of the 17th ACM conference on Information and In Proceedings of the Twelfth ACM International Conference
knowledge management, ACM: Napa Valley, California, on Web Search and Data Mining, pp. 717-725, 2019.
USA., pp. 931-940, 2008. [26] S. Rafiee, C. Salavati, and A. Abdollahpouri, "CNDP:
[11] A.J.B. Chaney, D.M. Blei, and T. Eliassi-Rad, "A Link prediction based on common neighbors degree
Probabilistic Model for Using Social Networks in penalization," Physica A: Statistical Mechanics and its
Personalized Item Recommendation," in Proceedings of the Applications, 539, 2020.
9th ACM Conference on Recommender Systems, ACM: [27] Y. Zhu, D. Huang, W. Xu, and B. Zhang, "Link prediction
Vienna, Austria, pp. 43-50, 2015. combining network structure and topic distribution in large-
[12] Y. Chen, and J. Canny, "Probabilistic recommendation of scale directed network," Journal of Organizational Computing
an item," Google Patents, 2011. and Electronic Commerce: 1-17, 2020.
[13] H. Ma, I. King, M.R. Lyu, Learning to recommend with [28] J. Liu, X. Jin, Y. Hong, F. Liu, Q. Chen, Y. Huang, and
social trust ensemble, in Proceedings of the 32nd international F. Sun, "Collaborative linear manifold learning for link
ACM SIGIR conference on Research and development in prediction in heterogeneous networks," Information
information retrieval, ACM: Boston, MA, USA., pp. 203-210, Sciences, 511, p.p. 297-308, 2020.
2009. [29] F. Maazouzi, H. Zarzour, and Y. Jararweh, "An effective
[14] Y. Zhang, M. Chen, D. Huang, D. Wu, and Y. Li, recommender system based on clustering technique for ted
"iDoctor: Personalized and professionalized medical talks," International Journal of Information Technology and
recommendations based on hybrid matrix factorization," Web Engineering (IJITWE) 15, no. 1, p.p. 35-51, 2020.
Future Generation Computer Systems, Vol. 66, pp. 30-35, [30] X. Ma, D. Guo, L. Cui, X. Li, X. Jiang, and X. Chen,
2017. "SOM Clustering Collaborative Filtering Algorithm Based on
[15] M. Jamali, and M. Ester, "A matrix factorization Singular Value Decomposition," In Proceedings of the 2019
technique with trust propagation for recommendation in social 4th International Conference on Mathematics and Artificial
networks," in Proceedings of the fourth ACM conference on Intelligence, p.p. 61-65. 2019.
Recommender systems, ACM: Barcelona, Spain, pp. 135-142, [31] H. Parvin, P. Moradi, and S. Esmaeili, "TCFACO: Trust-
2010. aware collaborative filtering method based on ant colony
[16] H. Liu, F. Xia, Z. Chen, N.Y. Asabere, J. Ma, and R. optimization," Expert Systems with Applications, 118, p.p.
Huang, "TruCom: Exploiting Domain-Specific Trust 152-168, 2019.
Networks for Multicategory Item Recommendation," IEEE [32] N. Tohidi, and C. Dadkhah, "Improving the performance
Systems Journal, Vol. 11, No. 1, pp. 295-304, 2017. of video Collaborative Filtering Recommender Systems using
[17] G. Guo, J. Zhang, and N. Yorke-Smith, "TrustSVD: Optimization Algorithm," International Journal of Nonlinear
collaborative filtering with both the explicit and implicit Analysis and Applications, 11(1), p.p. 283-295, 2020.
influence of user trust and of item ratings," in Proceedings of
M. Khalaji, C.Dadkhah & J.Gharibshah: Hybrid movie recommender system based on resource allocation (Regular Paper) 24

[33] N. Khodaverdi, C. Dadkhah, and M. Khalaji, "Movie [48] T.D. Pessemier, J. Dhondt, and L. Martens, "Hybrid
Hybrid Recommender System Based on Clustering and group recommendations for a travel service," Multimedia
Popularity," International Conference on Technology and Tools and Applications, Vol. 76, No. 2, pp. 2787-2811, 2017.
Innovation in Science, Engineering and Technology (TIET), [49] S. Wei, X. Zheng, D. Chen, and C. Chen, "A hybrid
2019. approach for movie recommendation via tags and ratings,"
[34] M. Khalaji, "NWS_RS: Personalizing Electronic Commerce Research and Applications, Vol. 18, pp.
Recommendation Based on Users Information and 83-94, 2016.
[50] Y. Deldjoo, M. Schedl, B. Hidasi, and P. Knees,
New Weighted Similarity Measure," Research in "Multimedia recommender systems," In Proceedings of the
Computer Science (IRAN), Persian Format, Vol. 13, pp. 1-14, 12th ACM Conference on Recommender Systems, pp. 537-
2019. 538, 2018.
[35] M. Khalaji, and N. Mohammadnejad, "CUPCF: [51] J.K. Tarus, Z. Niu, and D. Kalui, "A hybrid recommender
combining users preferences in collaborative filtering for system for e-learning based on context awareness and
better recommendation," SN Appl. Sci. 1: 1053, 2019. sequential pattern mining," Soft Computing, Vol. 22, No. 8,
https://doi.org/10.1007/s42452-019-1071-6 pp. 2449-2461, 2018.
[36] R.J. Mooney, and L. Roy, "Content-based book [52] J.A. Konstan, B.N. Miller, D. Maltz, J.L. Herlocker, .L.R.
recommending using learning for text categorization," in Gordon, and J. Riedl, "GroupLens: applying collaborative
Proceedings of the fifth ACM conference on Digital libraries, filtering to Usenet news," Commun, ACM, Vol. 40, No. 3, pp.
ACM: San Antonio, Texas, USA., pp. 195-204, 2000. 77-87, 1997.
[37] Y. Deldjoo, M. Elahi, P. Cremonesi, F. Garzotto, P. [53] J.L. Herlocker, J.A. Konstan, L.G. Terveen, and J.T
Piazzolla, and M. Quadrana, "Content-Based Video Riedl, "Evaluating collaborative filtering recommender
Recommendation System Based on Stylistic Visual Features," systems," ACM Trans. Inf. Syst., Vol. 22, No. 1, pp. 5-53,
Journal on Data Semantics, Vol. 5, No. 2, p. 99-113, 2016. 2004.
[38] R.V.D. Oord, S. Dieleman, and B. Schrauwen, "Deep [54] http://lectureonline.cl.msu.edu/~mmp/labs/error/e2.htm
content-based music recommendation," in Proceedings of the [55] B.K. Patra, R. Launonen, V. Ollikainen, and S. Nandi,
26th International Conference on Neural Information “A new similarity measure using Bhattacharyya coefficient for
Processing Systems, Vol. 2, Curran Associates Inc.: Lake collaborative filtering in sparse data,” Knowl.-Based Syst. 82
Tahoe, Nevada, pp. 2643-2651, 2013. (C), pp. 163–177, 2015.
[39] J. Yang, H. Wang, Z. Lv, W. Wei, and H. Song,
"Multimedia recommendation and transmission system based
on cloud platform," Future Generation Computer Systems,
Vol. 70, pp. 94-103, 2017.
[40] D. Wang, Y. Liang, D. Xu, X. Feng, and R. Guan, "A Mostafa Khalaji received the B.Sc.
content-based recommender system for computer science degree in computer engineering
publications," Knowledge-Based Systems, 157, pp.1-9, 2018. (Software) from Sadra Institute of
[41] B.R. Cami, H. Hassanpour, and H. Mashayekhi, "User Higher Education, Tehran, Iran, in 2015,
preferences modeling using dirichlet process mixture model and the M.Sc. degree in computer
for a content-based recommender system," Knowledge-Based engineering (Artificial Intelligence)
Systems, Vol. 163, pp. 644-655, 2019. from K. N. Toosi University of
[42] M. Lee, P. Choi, and Y. Woo, "A Hybrid Recommender Technology, Tehran, Iran, in 2017. He is
System Combining Collaborative Filtering with Neural a lecturer at Islamic Azad University,
Network," Berlin, Heidelberg: Springer Berlin Heidelberg, Shahr-e-Qods Branch and Sadra Institute of Higher
2002. Education. He is a member of IEEE and IAENG. His current
[43] S. Nadi, M. Saraee, A. Bagheri, and M. Davarpanh Jazi, research interests include Recommender Systems, Machine
"FARS: Fuzzy Ant based Recommender System for Web Learning, Social Network Analysis, and Data Mining.
Users," International Journal of Computer Science Issues, Email: Khalaji@email.kntu.ac.ir
Vol. 8, No. 1, pp. 203-209, 2011.
[44] T.H. Roh, K.J. Oh, and I. Han, "The collaborative Chitra Dadkhah received the B.Sc. degree
filtering recommendation based on SOM cluster-indexing in software engineering from Shahid
CBR," Expert Systems with Applications, Vol. 25, No. 3, pp. Beheshti University, Tehran, Iran, in 1990,
413-423, 2003. the M. Sc. degree in computer engineering
[45] J. Mai, Y. Fan, and Y. Shen, "A Neural Networks-Based (Artificial Intelligence) from the IASI
Clustering Collaborative Filtering Algorithm in E-Commerce Department, University of Paris 11, Orsay,
Recommendation System," in 2009 International Conference France, 1993. She also received the Ph.D.
on Web Information Systems and Mining, 2009. degree in computer engineering from the
[46] D. Kim, C. Park, J. Oh, and H. Yu, "Deep hybrid Department of Computer Engineering & IT,
recommender systems via exploiting document context and Amir Kabir University (Poly technique), Tehran, Iran, 2005.
statistics of items," Information Sciences, 417, pp. 72-87, She is an Assistant Professor in the Computer Engineering
2017. Faculty of K. N. Toosi University of Technology, Tehran,
[47] R. Katarya, and O.P. Verma, "Efficient music Iran. Her research interests include Evolutionary & Swarm
recommender system using context graph and particle Algorithms, Natural language processing, Recommender
swarm," Multimedia Tools and Applications, Vol. 77, No. 2, Systems, and Robotics (simulation).
pp. 2673-2687, 2018. Email: dadkhah@kntu.ac.ir
The CSI Journal on Computer Science and Engineering, Vol. 17, No. 2, 2020 25

Joobin Gharibshah received the B.Sc.

degree in software engineering, the master’s
degree in computer science, and the Ph.D.
degree in computer science from the
University of California Riverside, in 2020,
USA.
His research interests include
recommendation systems, search and ranking methods,
natural language processing, and machine learning. He is an
applied researcher at Ebay Inc.
Email: jghar002@ucr.edu

Paper Handling Data:

Submitted: 10.06.2019
Received in revised form: 10.28.2020
Accepted: 11.20.2020
Corresponding author: Dr. Chitra Dadkhah
Affiliation of the corresponding author: Faculty of
Computer Engineering, K. N. Toosi University of
Technology, Tehran, Iran