You are on page 1of 6

2018 9th International Conference on Information and Communication Systems (ICICS)

Improving Collaborative Filtering Recommender


Systems Using Semantic Information
Bushra Alhijawi*1 , Nadim Obeid+1 2 , Arafat Awajan†1 , Sara Tedmori◦1
1 King
Hussien School of Computer Sciences, Princess Sumaya University for Technology Amman, Jordan
2 King Abdullah II of Information Technology, The University of Jordan Amman, Jordan

Email: bus20179001@std.psut.edu.jo, + N.obeid@psut.edu.jo, † awajan@psut.edu.jo, ◦ S.Tedmori@psut.edu.jo


*

Abstract—The virtual world is overflowing with digital items similar users have liked without any understanding of the item
which makes user’s experience when searching for, selecting and itself. In other words, if user X and user Y gave rates to
buying items harder. The use of a recommender system can help n items in a similar way then there is a high probability
alleviate this problem and increase the profit of companies by
generating to the user a list of potential favorite items. This that they will give similar rates to other items [11]. CF can
paper proposes a new hybrid algorithm that depends on the be classified as: memory-based CF or model-based CF [11].
item’s semantic information and the users historical rating data. Memory-based CF uses the historical users rating data in order
The semantic information is used to estimate the strength of to calculate similarity values that are used to generate the
the semantic similarity between users in terms of preferred and recommendation [12]. The model-based CF uses the historical
non-preferred items. The historical rating data is used to estimate
the similarity in satisfaction level between users. Based on both users rating data to create a model such as Bayesian classifier,
similarity estimates, a unified list of neighbors is produced. This fuzzy algorithm, clustering algorithm and genetic algorithm
list will be used in the prediction step. in order to generate the recommendation [13]. Demographic-
Recommendation results obtained from employing the pro- based and knowledge-based are two additional recommenation
posed hybrid algorithm have been compared against recommen- techniques. The former is based on demographic profile of the
dation results from alternative techniques. The results obtained
show that the proposed algorithm outperforms the baseline user (e.g age and gender) and used to predict different items
techniques in recommendation quality and prediction accuracy. for different demographic statuses, while the latter is based
on inferences about a user’s needs and preferences. Finally,
I. I NTRODUCTION a hybrid recommender system is used to refer to any system
To help reduce problems associated with information over- that combines two or more recommendation techniques.
load, the use of Recommender Systems (RSs) has been The full dependency on the historical users-item rating
common in applications that offer users large amounts of matrix to generate recommendation leads to two problems
digital items. A RS that provides users with an accurate list that affect the recommendation quality: sparsity and cold-start
of recommendations (i.e. potential favorite items or interests problems. The sparsity problem is very common in CF [14],
of other pertinent data) will both attract users and incur [15]. This problem can be attributed to the fact that most users
benefits to companies [1]–[3]. It has various applications in e- do not rate most of the items or only rate a few. Hence, the
business (i.e. e-government, e-commerce/e-shopping, e-library, available ratings are sparse. This problem results in user-item
e-learning, e-tourism, e-resource services, e-group activities) rating matrix that contains a few nonzero entries. Cold start
[3], the DVD rental provider Netflix [4] and the online book problem arises when there is not enough information available
retailer Amazon.com [5]. Moreover, social networks such to the RS about the Active User (AU) as he/she is a new user or
as Youtube [6] and Facebook [7], mobile applications and when a new item is introduced [14], [16], [17]. This problem
services [8] and cloud computing services [9] harness the makes accurate prediction by the RS harder. A CF requires
power of RSs. large volume of existing user data in order to make accurate
A RS is a filtering system that tries to predict the rating predictions. This large volume of data is at a constant increase
or preferences a user would give to a digital item or to a as new ratings are made and as new users and items are added
social element (e.g. people) that he/she has not yet considered leading to the renowned scalability problem.
[10]. Content-based (CB) and Collaborative Filtering (CF) In this research, the authors aim to improve the recom-
are the two traditional and widely employed recommendation mendation’s quality of the CF and overcome the cold-start
techniques. CB filtering techniques generate their prediction and sparsity problems. The proposed technique exploits the
using a model that analyses the description of the item semantic information about the items to handle the cold-start
and the user’s preferences. In CB filtering techniques, the problem of CF and the historical users-item rating matrix to
recommended items are usually similar to items the user alleviate the overspecialization problem of knowledge-based
liked in the past. CF methods on the other hand generate RS. First, the Semantic Similarity (SemSim) between the AU
their prediction using a model that analyses large amounts of and other users in terms of items is calculated in two different
information relating to user’s historical rates. In CF methods, aspects: the preferred and non-preferred items of AU and
the items recommended to a user are usually similar to what other users. Based on the SemSim the most N similar users

978-1-5386-4366-2/18/$31.00 ©2018 IEEE 127


2018 9th International Conference on Information and Communication Systems (ICICS)
are selected as semantic neighbors. At the same time, the the filtering criteria used. The semantic information is used
normalized Pearson Similarity (PerSim) is computed between to handle the cold-start and sparsity problems. The author
AU and other users to compute the similarity in term of evaluated the system using MovieLens and synthetic dataset.
satisfaction degree (i.e. depends on the historical rates). Then, The results show that the use of semantic information enhances
the users who are satisfied and dissatisfied for quite similar the accuracy of CF by handling the cold-start and sparsity
particular items as the AU are selected to produce another list problems.
of neighbors. Depending on both lists of neighbors, a unified In the proposed RS, the genre, actors, director and origin
list of neighbors is produced and then used to generate the features of the movie are taken into its account to compute
recommendation. the SemSim between the AU and other users in term of items.
The rest of this paper is organized as follows. Section II Two SemSim measures are calculated:
provides an overview of some related worked. Section III 1) The SemSim between the AU and a user u based on the
describes the details of recommendation generation process. items that they gave a rate at least 3 (i.e. The SemSim
Section IV focuses on the testing of the algorithm developed in between the favorite items of AU and the favorite items
this work and shows the comparison results with the traditional of u).
CF techniques. Finally, section V concludes the paper and 2) The SemSim between the AU and a user u based on the
presents avenues for future work. items that they gave a rate at most 2 (i.e. The SemSim
between the items that AU is disliked it and the items
II. R ELATED W ORKS
that u is disliked it).
The semantic information has been utilized in many RS
At the same time, the satisfaction-based similarity (PerSim)
applications such as e-learning [18]–[22], movie recommen-
between the AU and other users is computed. Then, both
dation [23]–[26] and tourism recommendation [27], [28]. This
SemSim and PerSim are used to predict the rates of unrated
research focuses on hybrid-RS for movie that combines the
movies. Next, the proposed technique will be detailed.
CF with semantic information to improve the results of CF.
Lv et al. [23] developed a recommendation technique that III. R ECOMMENDATION G ENERATOR
combines genetic-based RS and domain ontology to overcome
In this section, the technique proposed to recommend the
the cold-start problem. Their system generates the recommen-
AU desired digital items depending on the SemSim and
dation depending on the item’s weighted semantic features.
PerSim is detailed. First, how to calculate the SemSim and
The weights of the features are obtained using the genetic
PerSim are described. After that, the technique of selecting
algorithm. For testing, the authors used the MovieLens data.
a unified list of AU’s neighbors is explained. Finally, the
Nilashi et al. [24] presented a new hybrid-RS that combines
recommendation is generated to the AU depending on the
CF, ontology and Singular Value Decomposition (SVD). The
selected list of neighbors.
domain ontology and SVD are adopted to solve sparsity and
For each user and based on the user-item rating matrix, the
scalability problem, respectively. For testing, the authors used
items are classified into two groups: preferred items (LikeSet)
MovieLens and Yahoo! Webscope R4 datasets. Cheng et al.
and non-preferred items (DisLikeSet). The LikeSet consists
[26] developed an ontology-based personalized RS to handle
of all items which the user u gave it a rate at least 3.
cold-start, scalability and suggestion-ability static problems.
The DisLikeSet consists of all items which the user u
Depending on the historical rating data and the semantic fea-
gave a maximum rate of 2. Then, the SimSem between the
tures of the highly rated items, the RS predicts the degree level
[LikeSetAU and LikeSetu ] and between the [DisLikeSetAU
the user might like an item. The authors used genetic algorithm
and DisLikeSetu ] are computed. These values indicate two
to train a personalized estimating module for each user that is
varying measures between AU and u as follows:
used later to generate the recommendation. Shambour and Lu
[29] developed a RS that improves the CF accuracy using the 1) The semantic similarity degree in relation to the items
semantic features of the items and user’s social trust network. they preferred. Thus, if AU and u have a high semantic
Their trust semantic fusion-based recommendation technique similarity in term of LikeSet then there is a high
handles both sparsity and cold-start problems. The evaluation probability that AU will prefer what u preferred and
of the proposed RS is done using Yahoo! Webscope R4 AU has not yet considered.
and MovieLens. Results obtained demonstrates the method’s 2) The semantic similarity degree in relation to the items
effectiveness in solving the mentioned problems. A hybrid they do not prefer. Thus, if AU and u have a high
Multi-Criteria Semantic-enhanced CF RS that combines CF semantic similarity in term of DisLikeSet then there
and semantic filtering technique is proposed in [12]. The is a high probability that AU will not prefer what u has
presented system addresses the sparsity and cold-start item not preferred and AU has not yet considered. Therefore,
problems. The authors evaluated their technique using Movie- those items should not be recommended to AU in the
Lens data. The experimental results of their work show that future.
the use of semantic information enhances the accuracy of Using the semantic information about the items to compute
multi-criteria CF. Alhijawi [30] presented a genetic-based RS the similarity will improve the accuracy of CF by decreasing
that uses the semantic similarity between the items as one of the effects of cold-start and sparsity problems. But this may

128
2018 9th International Conference on Information and Communication Systems (ICICS)
affect the accuracy because of overspecialization problem. – Compute the SemSim between the AU and u based
Therefore, the RS will not generate recommendations depend- on Eq. 6.
ing only on SemSim but also on PreSim (satisfaction-based). SemSim(AU, u) =
After computing the PerSim between the AU and other users, SemSimLike(AU, u) + SemSimDisLike(AU, u)
a unified list of neighbors is selected. This list will be used 2
to make a more accurate prediction and generate high-quality (6)
recommendations. P
SemSimI(IAU , Iu )
SemSimLike(AU, u) = ,
A. The Semantic-based Similarity |LikeSetAU | ∗ |LikeSetu |
IAU ∈ LikeSetAU and Iu ∈ LikeSetu
The SemSim between the users is dependent on the item- (7)
based SemSim. The item-based SemSim of a given item
against other items is computed based on the common se- SemSimDisLike(AU, u) =
mantic features. Each item is represented by a vector (Vi ) of P
SemSimI(IAU , Iu )
concepts and each concept is represented by a vector (Vc ). , (8)
|DisLikeSetAU | ∗ |DisLikeSetu |
This research takes into account four item features: genre, IAU ∈ DisLikeSetAU and Iu ∈ DisLikeSetu
actor, director and origin. Both genre and actor are multi-
valued attributes whereas origin country is a mono-valued
attribute. To do this, binary Jaccard similarity coefficient is B. The Satisfaction-based Similarity
used to measure the degree of overlap between the two item’s The P reSim is computed based on the items that are rated
features. The SemSim measure calculation steps are mentioned by both AU and u. Therefore, it is reflected the similarity
bellow: based on the satisfaction degree. But considering two users
• Each item i is represented by a vector as follows: are similar is inequitable when one of them rated a small
number of items and another user rated a large number of
Vi = (Vc1 , Vc2 , Vc3 , Vc4 ) (1) items. For instance, consider the case in which user A rated
20 items, user B rated 150 items and both users gave those
20 items same ratings. Then, it is inequitable to say that user
• Classify the items into two groups: preferred items A is 100% similar to user B because user A rated 13.4%
(LikeSet) and non-preferred items (DisLikeSet). of the items that are rated by user B. Therefore, the PerSim
• Compute the SemSim using the Jaccard similarity metric is normalized by the binary Jaccard similarity to get a more
(Eq. 2). accurate and reliable similarity value. The PreSim calculation
steps are detailed bellow:
F11 • Each user u is represented by a binary vector as follows:
SemSimI(p, q) = (2)
F10 + F01 + F11
Vu = (vui1 , vui2 , . . . , vuig ) (9)
Where,
Where,
– Vp = (Vpc1 , Vpc2 , Vpc3 , Vpc4 ).
– k ∈ {1, 2, 3, . . . , g} and g is the total number of
– Vq = (Vqc1 , Vqc2 , Vqc3 , Vqc4 ).
available items.
– F11 : Total number of features which both items, p
– In case vuik = 1, the user rated the item k. Otherwise,
and q, are fall into it. Formally,
vuik = 0.
4
X • Compute the percentage of common items between the
F11 = |Vpcn ∧ Vqcn | (3) AU and u using the binary Jaccard (Eq. 10).
n=1
|UAU ∩ Uu |
Jaccard(AU, u) = (10)
– F10 : Total number of features where p is fall into it |UAU ∪ Uu |
and q is not fall into it. Formally,
• Compute the Pearson correlation between AU and u using
4
X Eq. 11.
F10 = |Vpcn − Vqcn | (4)
n=1 P
− rAU
i∈I (rAU,i ¯ )(ru,i − r¯u )
P earson(AU, u) = qP qP
– F01 : Total number of features where q is fall into it i∈I (rAU,i − rAU
¯ ) 2
i∈I (ru,i − r¯u )
2

and p is no fall into it. Formally, (11)


Where,
4
– I is the group of items that both users AU and u
X
F01 = |Vqcn − Vpcn | (5)
n=1 have rated.
– rAU,i is the rate of user AU on item i.

129
2018 9th International Conference on Information and Communication Systems (ICICS)
– rAU
¯ is the mean rating value of user AU . Where,
– ru,i is the rate of user u on item i. • U is a set of users who are belong to UNL.
– r¯u is the mean rating value of user u. • r̄AU is the mean rating value of user AU .
• Compute the normalized similarity between AU and u • r̄u is the mean rating value of user u.
(P reSim(AU, u)) using Eq. 12.
IV. E VALUATION AND R ESULTS
P reSim(AU, u) = P earson(AU, u) ∗ Jaccard(AU, u)
(12) This section provides detailed of how the proposed hybrid
RS was tested. Various experiments were conducted for pur-
C. Selecting Neighbors poses of comparing the proposed algorithm with alternative
In this step, two lists of neighbors are produced. One list is techniques; specifically with cosine similarity and Pearson
a result from SemSim (i.e. SN L) and other is a result from correlation. Details related to the data that has been used in
PreSim (i.e. P N L). P N L contains the most N umN P similar the experiments is provided in subsection IV-A. Subsection
users to the AU based on satisfaction degree. In other words, IV-B provides details related to the experiments design and
P N L includes the users who are satisfied and dissatisfied the measures that were used in the experiments. Finally, the
for quite similar particular items as AU. But, these users do results are presented and discussed in subsection IV-C.
not necessarily have similar semantic behavior to AU . For
instance, user A and user B both gave the rate of 4 to a A. Experimental Data
particular movie. User A liked the movie because it is an In the experiments, HetRec 2011 (MovieLens +
action and adventure movie while user B enjoyed it since the IMDb/Rotten Tomatoes) dataset 1 has been used. This
actor act1 acted in this movie. SN L contains the most similar dataset consists of 855598 rating (1-5) collected from 2113
users to the AU semantically. However, the users within this users on 10197 movies, 95321 actors, 4060 directors and
list may have not the same satisfaction degree as AU for quite 20 genres. It is an extension of the MovieLens10M dataset,
similar particular items. Due to this, those lists are exploited published by GroupLeans research group 2 . Table I shows the
to produce a more accurate list (i.e. Unified Neighbors List statistics of the experimental data.
(UNL)) which contains the users who are similar to the AU
B. Experiments Design
semantically and in satisfaction degree. This list is generated
using the Eq. 13 which depends on both lists: P N L and SN L. Two different experiments were conducted. The first exper-
UNL will be used to make a more accurate prediction and iment aims to compare the proposed RS with the traditional
generate a high-quality recommendation. Figure 1 illustrate CF techniques (i.e. Pearson correlation-based CF and cosine-
the process. based CF). The second experiment aims to justify the multi-
plication operation between Jaccard and Pearson values when
computing the PreSim. In order to perform this experiment,
the PreSim is calculated using two equations; the first equation
is essentially the product Jaccard and Pearson similarity and
the second equation is using only Pearson similarity. Each RS
ran 250 times and the average of the 250 runs was taken to
get more accurate results.
The results of the proposed RS are compared with the
ones obtained from the other approaches using Mean Absolute
Error (MAE) and the recommendation quality metrics: recall,
Fig. 1. The process of producing the unified neighbors lists. precision and F1-measure. MAE (Eq. 16) is a quantity used to
measure how close prediction rates are to the real rates. The
U N LV = |P N L ∩ SN L| (13) prediction accuracy is computed using a different number of
neighbor users for the AU;(K). The value of K range between
20 to 140. Recall (Eq. 17) represents the percentage of the
D. Producing recommendation system sensitivity or the system ability to gather the favorite
items to the AU. While precision (Eq. 18) represents the
The selected neighbors (i.e. UNL) are used to predict the
system ability to suggest items that are truly interesting for the
rates of unrated items by the AU. The prediction rates are
AU. F1-measure (Eq. 19) is the harmonic mean of precision
calculated using Eq. 15. Then, The T opN items with the
and recall. Those metrics are calculated using different values
highest prediction rates are captured and recommended to the
of N (number of recommendations made) ranging from 10
AU.
to 100. All users and items in the data set are considered
as historical data except one user (u) which is considered
Sim(AU, u) = P reSim(AU, u) + SemSim(AU, u) (14)
as an AU. In each run, AU is selected randomly. For testing
∗ (rui − r̄u )]
P 1 https://grouplens.org/datasets/hetrec-2011/
u∈U [Sim(AU, u)
piAU = r̄AU + P (15) 2 https://grouplens.org/
u∈U Sim(AU, u)

130
2018 9th International Conference on Information and Communication Systems (ICICS)
TABLE I
DATA S TATISTICS

Ratings/User Ratings/Movie Origin country/Movie Actors/Movie Director/Movie Genres/Movie


404.921 84.637 1 22.778 1 2.040

purposes, 50% of the items which were rated by u hid and make the prediction. On average, the results obtained from
assuming that the proposed RS should recommend to u a set using Method 1 are 13%, 23% and 29.6% better than Method
of items which belong to the hided items set and its ratings 2, 3 and 4, respectively.
are 3, 4 or 5.
U PIu
1 X i=1 |pu,i − ru,i |
M AE = (16)
#U u=1 #Iu

TP
Recall = (17)
TP + FN

TP
P recision = (18)
TP + FP

2 ∗ P recision ∗ Recall
F 1 − measure = (19)
P recision + Recall
C. Results Fig. 3. The comparison between Method 1 and other methods based on the
precision.
In this section, the results obtained from all experiments
are presented and discussed. The objective is to compare the
proposed technique to the alternative techniques in terms of
prediction accuracy and recommendation quality. Note that
Method 1, 2, 3 and 4 are the proposed technique, the proposed
technique without normalizing the Pearson similarity, Pear-
son correlation-based technique and cosine-based technique,
respectively.

Fig. 4. The comparison between Method 1 and other methods based on the
F1-measure.

Figs. 3 and 4 show the precision and F1-measure results,


respectively. As known, precision and F1-measure reach its
best value at 1 and worst at 0. The precision of all methods
decreases when the number of recommendations increases. As
Fig. 2. The comparison between Method 1 and other methods based on the
shown, the precision of Method 1 is close to the recommenda-
accuracy. tion quality of Method 2. Moreover, the precision of Method
2 is better than that of Method 1 when N = {60, 70, 90}.
Fig. 2 shows the MAE results obtained from all methods. However, Method 1 outperforms Method 2, 3 and 4 by 1.6%,
As may be seen in this figure, the accuracy degree of Method 78.2% and 161.7%, respectively. This indicates that Method 1
1 is far better than the results of Method 2, 3 and 4 for any can offer more reliable and interesting items to users than
selected value of the number of neighbors (k). The accuracy other methods. Based on Fig. 4, one can observe that the
of all methods is improved when the k is increased. But the recommendation quality of Method 1 is much higher than the
MAE results of Method 1 when k > 80 are so close for recommendation quality of the other methods. Also, it can be
any selected value of k. This means that Method 1 is able clearly seen that there is a positive correlation between the
to predict the rates accurately without the need for a large F1-measure and N . This result indicates that the recall results
number of neighbors. Note that selecting a larger number of of Method 1 outperforms the recall of the other methods. This
neighbors increases the computational time that is needed to is due to the fact that F1-measure depends on both recall and

131
2018 9th International Conference on Information and Communication Systems (ICICS)
precision. The F1-measure results of Method 1 are better than [8] J. Zeng, F. Li, H. Liu, J. Wen, and S. Hirokawa. A restaurant
Method 2,3 and 4 by 0.86%, 90.4% and 170.7%, respectively. recommender system based on user preference and location in mobile
environment. In 5th IIAI International Congress on Advanced Applied
The second experiment aims to justify the multiplication op- Informatics (IIAI-AAI), pages 55–60. IEEE, 2016.
eration between Jaccard and Pearson values when computing [9] D. Leony, H. Gelvez, A. Munoz-Merino, A.Pardo, and C. Kloos. A
the PreSim. As mentioned, in this experiment the PreSim is generic architecture for emotion-based recommender systems in cloud
learning environments. Journal of Learning and Collaboration Tech-
calculated using two equations; the first by multiply Jaccard nologies: Technology-Rich Environments for Learning and Collabora-
and Pearson similarity (i.e. Method 1) and the second by only tion, LNCS, 8524:105–116, 2014.
Pearson similarity (i.e. Method 2). Figs. 16, 3 and 4 clearly [10] L. Sharma and A. Gera. A survey of recommendation system: Research
challenges. International Journal of Engineering Trends and Technology
show that the accuracy and recommendation quality of Method (IJETT), 4(5):1989–1992, 2013.
1 is better than those obtained from using Method 2. [11] X. Su and T. Khoshgoftaar. A survey of collaborative filtering tech-
niques. Advances in artificial intelligence, 2009:4, 2009.
V. C ONCLUSION [12] Q. Shambour and J. Lu. A hybrid multi-criteria semantic-enhanced
collaborative filtering approach for personalized recommendations. In
In this research, a new hybrid RS that improves the rec- the 2011 IEEE/WIC/ACM International Conferences on Web Intelligence
and Intelligent Agent Technology-Volume 01, pages 71–78. IEEE Com-
ommendation quality and prediction accuracy of CF is pro- puter Society, 2011.
posed. This RS uses the semantic information about the item [13] J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez. Recommender
to compute the semantic similarity between the users. This systems survey. Knowledge-based systems, 46:109–132, 2013.
[14] G. Guo, J. Zhang, and D. Thalmann. Merging trust in collaborative
improved the accuracy of CF by decreasing the effects of cold- filtering to alleviate data sparsity and cold start. Knowledge-Based
start problem. But, depending solely on the semantic similarity Systems, 57:57–68, 2014.
may affect the accuracy because of overspecialization problem. [15] F. Xie, Z. Chen, J. Shang, and G.C. Fox. Grey forecast model for
accurate recommendation in presence of data sparsity and correlation.
Therefore, the recommendation does not generate recommen- Knowledge-Based Systems, 69:179–190, 2014.
dations depending only on semantic similarity but also on [16] A.L.V. Pereira and H.R. Hruschka. Simultaneous co-clustering and
satisfaction-based similarity. In order to compute the semantic learning to address the cold start problem in recommender systems.
Knowledge-Based Systems, 82:11–19, 2015.
similarity, the rated items by each user are classified into two [17] K. Ji and H. Shen. Addressing cold-start: Scalable recommendation with
groups preferred items and non-preferred items. The semantic tags and keywords. Knowledge-Based Systems, 83:42–50, 2015.
similarity between the preferred items of AU and other users [18] M. Khobreh, F. Ansari, M. Dornhöfer, and M. Fathi. An ontology-based
recommender system to support nursing education and training. In LWA,
is computed to find the semantic similarity degree in term pages 237–244, 2013.
of what they preferred. The semantic similarity between the [19] B. Vesin, M. Ivanović, A. KlašNja-MilićEvić, and Z. Budimac. Protus
non-preferred items of AU and other users is computed to 2.0: Ontology-based semantic recommendation in programming tutoring
system. Expert Systems with Applications, 39(15):12229–12246, 2012.
find the semantic similarity degree in terms of what they do [20] L. Zhuhadar, O. Nasraoui, R. Wyatt, and E. Romero. Multi-model
not preferred. The satisfaction-based similarity is computed ontology-based hybrid recommender system in e-learning domain. In
using the Pearson correlation and depending on the historical IEEE/WIC/ACM International Joint Conferences on Web Intelligence
and Intelligent Agent Technologies, WI-IAT’09, volume 3, pages 91–95.
rating data. Next, a unified list of neighbors is produced based IEEE, 2009.
on both semantic and satisfaction-based similarity. This list [21] I. Ciuciu and Y. Tang. A personalized and collaborative elearning
includes the users who have similar semantic behavior and materials recommendation scenario using ontology-based data matching
strategies. In On the Move to Meaningful Internet Systems: OTM 2010
satisfaction degree to the AU . Workshops, pages 575–584. Springer, 2010.
The proposed RS is compared to the baseline techniques [22] J. K. Tarus, Z. Niu, and A. Yousif. A hybrid knowledge-based
in terms of prediction accuracy and recommendation quality. recommender system for e-learning based on ontology and sequential
pattern mining. Future Generation Computer Systems, 72:37–48, 2017.
The results obtained prove that the proposed algorithm has [23] G. Lv, C. Hu, and S. Chen. Research on recommender system based on
better performance in terms of accuracy and quality compared ontology and genetic algorithm. Neurocomputing, 187:92–97, 2016.
to the alternative technique. Also, the results show that the [24] M. Nilashi, O. Ibrahim, and K. Bagherifard. A recommender system
based on collaborative filtering using ontology and dimensionality
proposed technique is capable of making high-quality recom- reduction techniques. Expert Systems with Applications, 92:507–520,
mendation regardless of the number of selected neighbors and 2018.
recommendations. [25] I. Cantador, A. Bellogı́n, and P. Castells. A multilayer ontology-based
hybrid recommendation model. Ai Communications, 21(2-3):203–210,
2008.
R EFERENCES [26] S.-T. Cheng, C.-L. Chou, and G.-J. Horng. The adaptive ontology-based
[1] B. Alhijawi and Y. Kilani. Using genetic algorithms for measuring the personalized recommender system. Wireless personal communications,
similarity values between users in collaborative filtering recommender 72(4):1801–1826, 2013.
systems. In Computer and Information Science (ICIS), 2016 IEEE/ACIS [27] C.-I. Lee, T.-C. Hsia, H.-C. Hsu, and J.-Y. Lin. Ontology-based tourism
15th International Conference on, pages 1–6. IEEE, 2016. recommendation system. In 4th International Conference on Industrial
[2] X. Yang, Y. Guo, Y. Liua, and H. Steck. A survey of collaborative Engineering and Applications (ICIEA), pages 376–379. IEEE, 2017.
filtering based social recommender systems. Computer Communications, [28] Z. Bahramian and R. A. Abbaspour. An ontology-based tourism recom-
41:1–10, 2014. mender system based on spreading activation model. The International
[3] J. Lu, D. Wu, M. Mao, W. Wang, and G. Zhang. Recommender system Archives of Photogrammetry, Remote Sensing and Spatial Information
application developments: A survey. Decision Support Systems, 74:12– Sciences, 40(1):83, 2015.
32, 2015. [29] Q. Shambour and J. Lu. A trust-semantic fusion-based recommenda-
[4] Netflix. Date of access: 20/11/2017. tion approach for e-business applications. Decision Support Systems,
[5] Amazon. Date of access: 20/11/2017. 54(1):768–780, 2012.
[6] Youtube. Date of access: 20/11/2017. [30] B. Alhijawi. The use of the genetic algorithms in the recommender
[7] Facebook. Date of access: 20/11/2017. systems. Master’s thesis, Hashemite University, 2017.

132

You might also like