BULGARIAN ACADEMY OF SCIENCES

101

CYBERNETICS AND INFORMATION TECHNOLOGIES  Volume 19, No 1

Sofia  2019 Print ISSN: 1311-9702; Online ISSN: 1314-4081

DOI: 10.2478/cait-2019-0006

Hybrid Recommender System via Personalized Users’ Context

Anthony Nosshi

, Aziza Asem

, Mohamed Badr Senousy

Information System Dept., Computers and Information Faculty, Mansoura University, Mansoura, Egypt

Computer and Information System Department, Sadat Academy for Management Sciences, Cairo, Egypt

E-mails: [email protected] [email protected] [email protected]

Abstract: In movie domain, finding the appropriate movie to watch is a challenging

task. This paper proposes a recommender system that suggests movies in cinema that

fit the user’s available time, location, mood and emotions. Conducted experiments

for evaluation showed that the proposed method outperforms the other baselines.

Keywords: Movie recommender, Emotion recommendation, Hybrid recommender

system, sentiment analysis, spatio-temporal recommendation.

1. Introduction

Recommender systems help users overcome the information overload problem. They

can be classified into three main categories: Content-Based (CB), Collaborative

Filtering (CF) and Hybrid technique. In CB technique, the system builds a user profile

based on his/her preferences, then it finds the best fits to this profile [1]. It is helpful

in discovering different users’ interests [2]. In CF, the system uses other users’ ratings

to items to understand what the user would like or dislike [3]. The CF can give

recommendations to the user that is different from what the user had seen before [4].

Despite the benefits provided by these methods, they still suffer from many

issues. For example, CF suffers from sparsity problem, which causes the

recommender system to give poor recommendations [5]. Moreover, the CF also

suffers from the cold start problem [6]. CB methods also suffer from problems such

as the “over-specialization” problem. This problem means the system recommends

similar items to what the user has seen [7].

Today’s websites and social media enable users to leave their feedback and

opinions to be read publicly [8]. Consequently, this enables researchers to extract the

users’ interests, mood or contextual information from the social media posts they

share publicly. Therefore, this information can be incorporated into a hybrid

recommender system to produce a better recommendation [9]. Using a hybrid

recommender system which applies both aforementioned recommendation

techniques allows incorporating additional sources of information and content to

produce better recommendations. Therefore, studying additional information about

102

the user helps in customizing the recommendation to every user according to his/her

needs [10]. Thus, information such as the user’s mood during the day, the time of

recommendation, and the location of the user can be very helpful in enhancing the

recommendation process [11].

This paper proposes a hybrid recommender system that filters the

recommendation list using three phases filtering approach. The recommender is

applied to the movies in the cinema’s domain. In the first phase, the system

determines the current user mood to find the most suitable recommendation. Then,

the system in the second phase filters the produced list according to the time and

location of the user to recommend him/her the nearest cinema that has the preferred

movie (Spatio-Temporal Factors). Finally, the system filters the produced list by

determining the emotional factors that match with the user’s emotions which he/she

got from watching another movie previously, regardless of its genre. The remainder

of this paper is organized as follows: Section 2 discusses the state of the art, Section

3 discusses the proposed recommender, Section 4 discusses the results and evaluation

and finally, Section 5 discusses the conclusion and future work.

2. State of the art

Many studies have been conducted by researchers to enhance the recommendation

results by incorporating the user’s produced contents, sentiments, locations, and other

features as follows:

Q. Y a n g [12] proposed a recommender system that integrated both the domain

semantics along with the context information. For that purpose, the author developed

an improved content-based model to incorporate them. Though this work took into

account only four types of emotions. Additionally, it did not take care of the

emotional impression that a user may be willing to have as a result of watching a

movie. In addition to ignoring the current user’s mood, user’s location, and user’s

time availability. D i x i t and J a i n [13] proposed a recommender system that took

into account the contextual information of three different categories in order to

prepare their data selection and construction. However, this work depended only on

three main categories: demographic, semantic and social context and ignored the

emotional effect of watching movies. In addition, it missed the spatio-temporal

factors and the current user mood. L u i s et al. [14] proposed a system based on

semantic web technology. The system considered the location, time and the crowd of

people at the place of interest. Though, this work missed taking into consideration

the user’s mood. They depend on their work on location and time and the place crowd

only. A b b a s et al. [15] proposed a hybrid movie recommender system that took

into consideration the different users’ interests and then provided a recommendation

based on the context the users were in. Additionally, they applied their approach to

four different contexts and then compared how the system performed with these

different four contexts. C a i and Gu [16] proposed a recommender system using a

tensor factorization method. Though, their work missed the current user mood, the

spatio-temporal factors, and the emotional effect of the movies. Kim et al. [17]

proposed a recommender system that integrated convolutional neural network into

103

probabilistic matrix factorization. Their approach acquired the contextual

information from the document by applying the convolutional neural network to

enhance the accuracy of rating prediction. Though, they ignored the spatio-temporal

effects on the recommendation, the user mood, and the emotional factors as well. Y u,

L i n and W a n g [18] proposed a recommendation framework to alleviate the

sparsity problem. To this end, they built a contextual profile for each contextual

condition using a co-clustering algorithm. Additionally, they used the expanded

preferences in their recommendation system. Though, they ignored the emotional

effects on movies and matching the personal user’s emotions with the movie

emotions. Additionally, they missed the user mood. D e n g et al. [19] proposed a

recommender system that took into account the user’s emotions in different

granularity levels and different time windows. In comparison to the proposed work,

their work took into consideration only the emotional factors and time. They did not

take in consideration the spatial factors. Z h o u et al. [20] proposed a recommender

system that analyzed the movie-poster image and the text description for movies, user

ratings, and social relationship. Then, they utilized a random-walk methodology for

presenting the recommendation. Though, they ignored the spatio-temporal factors.

In summary, the aforementioned studies missed considering in their

recommendation process some or all of the following features: the current user mood,

the location and time of the recommendation, and the emotional factors in the time

domain.

3. The proposed recommender

The proposed algorithm consists of the three main phases, the first phase is

responsible for determining the user’s mood, the second is responsible for

determining the spatio-temporal factors and the final phase is responsible for the

emotional constraints as illustrated in Fig. 1.

Fig. 1. The proposed recommender overview

Start

Spatio-Temporal

Constraints (Lk)

Mood Determination (Lm)

End

Emotional Constraints

104

3.1. The first phase: Mood determination

The first step in the proposed recommender is to extract the user’s mode. In a nutshell,

this step consists of two subsystems as illustrated in Fig. 2. The first is (Data

Preparation) subsystem, the second is the (Mood Fitting) subsystem.

3.1.1. The first subsystem: Data preparation

The goal of this step is to extract the users’ emotions from both: The social media

posts and the shared movie posts that the user shared in a microblog publicly

(Fig. 2). Twitter was used for that purpose to extract users’ tweets. Then, some

associations between the users, emotions, and movies were constructed and

represented in a tuple of three elements (User, Emotion, Movie) as will come in more

details in the following paragraphs.

Fig. 2. Mood determinations subsystems

The idea behind the approach applied in this step is that there are associations

between the user posts on the social media and the movies they are watching [19].

The emotions expressed in the social media, just before the time that the user

expressed him/herself as watching a movie, can reflect to a great extent the mood the

user was in to take a decision of watching a specific movie. Applying the same

concept represented in [19], the social media posts can be classified into two types:

movie related posts, and general posts. The general posts are those, which users

usually express everything they want to, such as news, information, personal

opinions, etc. While movie posts are those, which the users share to show that, they

are watching a movie in their favorite cinema. Table 1 shows the difference between

the two types of social media posts. Both the first and the second posts belong to the

general post type, while the third belongs to the movie related post.

extracts

General Posts

Moods

Movies Posts

Users’

Social

Media

Movies

Total

Associations

(User, Mood,

Movie)

Target

User

Social

Media

General

Posts

Current

Mood

extracts

Association (User, Mood,

Movie)

Collaborative Filtering

List (Lm) of Movies fits

the user’s mood

A; Data Preparation

Subsystem

B: Mood Fitting

Subsystem

105

Table 1. Microblog example

USER ID

Post content

Time

Movie

876543467

Would like to thank all my friends for the birthday

wishes &…

2016.06.15

13:25:39

876543467

Feeling Excited with my friends

2016.06.15

13:29:12

876543467

Watching “Victoria”. Best birthday ever with my

lovely friends @Village East Cinema, 181-189 2nd

Ave, New York, NY 10003

2016.06.15

13:55:02

Victoria

The analysis of the above tweets can infer that the user before he/she has been

to the cinema was in a happy mood. He/she was celebrating with their friends by

watching a movie called “Victoria”. Then, by applying the aforementioned

hypothesis, it can be inferred that her decision to watch the movie with his/her friends

was according to the correlation between his/her happy mood and the movie they

watched. Thus, it was an emotional dependent decision. As illustrated in (Fig. 2), on

the first step in data preparation, Twitter posts (tweets) were extracted for people

located in the United States using Twitter Enterprise Data

(https://developer.twitter.com/en/enterprise). This extraction retrieved

60,345,567 users’ data with an average of 720 tweets per user. Then, users who

tweeted less than five movies or cinemas’ tweets were filtered out, also movies or

cinemas which were tweeted less than five times were removed. The dataset now

consists of a total of 12,927,849 tweets. Then we used SAS Viya

(https://www.sas.com/en_us/software/viya.html) to extract the sentiments

contained in those tweets as it has its own dictionary. Afterwards, all tweet posts were

segmented into words that can be used to construct the Bag Of Words model (BOW)

[21]. Then, based on that model, the number of emotional words were counted for

each emotional class [22] (Granularities of 2d, 7d, and 21d). According to the number

of emotional words, the emotional vector of the text was determined. Therefore, each

microblog post can be represented into emotional vectors. For example, the ﬁrst

microblog in Table 1 has three emotion vectors: 2D-emotion vector, 7D-emotion

vector, and 21D-emotion vector, the values of which are (3, 0), (3, 0, 0, 0, 0, 0, 0),

and (3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0). Then, to discover the

emotions and moods related to the movies’ posts, only the general social media posts

before posting about movies were taken into consideration. For that purpose, the

users’ latest social media posts were retrieved in the specified time window (last hour

in this case). Then, the sum of all qualified social media posts’ emotional vectors was

calculated. After that, associations were formed between the emotions, movies, and

all users. Then, the associations were represented by a three-element tuple (User,

Mood, Movie). Example (876,543,467, ((4,0), (3,1,0,0,0,0,0),

(3,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)), Victoria)

3.1.2. The second subsystem: Mood Fitting

This subsystem (Fig. 2, Mood Fitting) consists of three steps, the first step is for

extracting the current user’s associations, the second is for calculating the similarity,

and the third ‒ for finding out the current user’s interest.

106

 Step 1. The current user’s associations: All the target user’s posts (both

general and movie related posts) were extracted. Then, the user’s current mood and

the historical associations were also calculated by the same method discussed in the

previous subsystem.

 Step 2. Calculating the similarity: After that, the similarity between the

current user and the other users were computed depending on the associations they

have by incorporating it into collaborative filtering as presented in [19]. Any two

users were considered similar if they watched the same movie before. The similarity

degree increased with the more movies being common between the two users.

Additionally, if two users watched the same movie and they also shared the same

mood, they have more similarity than any other two users who watched the same

movie, but under different moods. For example, if three users A, B and C watched

the same movie and assuming that A and C watched it in a happy mood, and B

watched the same movie in an angry mood, then A and C have higher similarity

together than B. Then, the users were sorted according to their similarity to the target

user and we got the top-k user list.

 Step 3. The current user’s interest: Afterwards, the interest of the current user

to all movies of the top-k users was calculated as presented in [19] as follows: When

the top-k similar users watch a movie in the same current mood of the target user, it

is more likely that the user will have a higher interest in that movie. Then, all movies

were sorted according to the target user’s interest to them and added to a list (denoted

here as Lm) to be the input of the next step for spatio-temporal constraints phase.

3.2. The second phase: Spatio-temporal constraints

This phase receives the (Lm) list produced from the first phase (Mood Determination)

and then applies some processes to produce a filtered movie list that fits with spatio-

temporal constraints of the target user. It consists of three main components: User

Interface, User Profiler Engine, and Spatio-Temporal engine. It is responsible for

determining the spatio-temporal factors. They work as follows.

 User Interface. It is responsible for interacting with both the user and

the recommendation engine as well. It is built with Telegram Messenger

(Telegram Bot) (https://core.telegram.org/bots). It is also responsible

for determining the time and the location of the user using the Google Location

API (https://developers.google.com/location-context/fused-location-provider/)

(to retrieve the geographical coordinates), and Geocoding API

(https://developers.google.com/maps/documentation/geocoding/start) (for

reversing geocoding to get a human-readable address). Then, it sends them to the

User Profiler Engine to be used in the recommendation process. The time can be

either the current time that the user requested the recommendation at, or a time that

the user specified according to his/her wish.

 The User Profile Engine. Its job is to build a personal profile for the target

user in the system. The profile contains the target user’s mood extracted from the first

phase, and the retrieved movie list (Lm list), in addition to the time and location

retrieved from the user interface.

107

 Time Engine. This engine is responsible for filtering out the movie list

(Lm) to get the movies’ names that are running in cinemas in the user’s

preferred time. For that purpose, it first retrieves the user profile, which contains

the user’s preferred time. Then, it applies a programming interface that sends

HTTP requests to a movie’s API called the International Showtime API

(https://api.internationalshowtimes.com) to retrieve a list of movies that are in

cinemas at the user’s specified time range (date and hour). It also retrieves a list of

the cinemas displaying those movies with their details like the address, street, and

city. Afterwards, it filters the (Lm) list to pick up the movies and the cinemas in the

specified time range, i.e., it discards all the movies that are not in cinemas at the time

specified by the user. After filtering the (Lm) list, it creates a new filtered list for

movies in a suitable time (denoted here as Lt).

 Location Engine. It retrieves the lastly created list (Lt) that contains the

movies within the user’s preferred time. Then, it uses Google Distance Matrix API

(https://developers.google.com/maps/documentation/distance-matrix/start) to

retrieve the distance and time needed to reach to the cinemas that display those

movies. Therefore, it calculates the distance and the required time for the user to

move from the current location (the origin) and the destination cinemas, taking into

account the user’s mobility method (example: driving, walking, cycling, etc.).

Finally, it creates a new filtered list denoted here as (Lk). This list now contains the

movies that match the user mood, within her preferable time and in near cinemas to

his location. This list of movies (Lk) will be the input of the next third and final step

in the recommendation process, the EFBM step.

3.3. Exploring the emotional factors with emotional fingerprint model

Now, the prepared movies’ list (Lk) is passed to the third phase of the proposed

recommender in order to filter the movies for finding which movies can give the user

the desired emotions. For that purpose, the Emotional Fingerprints Based Model

(EFBM) proposed in [23] was applied. The EFBM model groups movies by

emotional patterns of some key factors that change across time, which were extracted

from the movies’ reviews. Those factors are what people discussed in reviews,

showing people’s most popular everyday problems, ideas, objects, and philosophical

concepts. These patterns in time form a kind of fingerprints that enable the user to

choose a movie that can give him/her the same emotional experience he/she got

before from watching and liking another movie in the past, regardless of its genre. To

apply the EFBM model, the IMDB (https://www.imdb.com/) users’ reviews were

first extracted for all movies which were included inside the (Lk) list. Then, these

customer reviews were analyzed to discover the movies’ features and make groups

for those movies depending on the features. Afterwards, the algorithm

was applied in Fig. 3.

108

Fig. 3. Algorithm used in the EFBM model [23]

In a nutshell, in the first step of the algorithm, two of small models were created

for the data dimension reduction and then aggregated later. After that, the genres were

combined with a technique of dimension expanding. The complementary weight was

used to modify the Jaccard distance, in order to calculate the value of the information

in a genres field. It was calculated as follows [23]:

(1) w

wgh

= –log(w

) * –log(w

where w

wgh

is the complementary weight; w

is the complementary weight of movie

by genre [23], and

(2) w



























where N is the movie count by all genres. Then, topics were extracted using SAS

Enterprise Miner (https://www.sas.com/en_us/home.html) to discover what every

cluster is about. After that, the features taxonomy was created manually. Then, the

sentiment analysis was applied to each cluster for visualization purposes.

Fig. 4. Statistics of sample clusters after applying sentiment analysis

109

Fig. 4 shows the statistics of sample clusters after sentiment analysis was applied

(A for cluster_id=100, B for cluster_id=143).

Then, SAX (http://www.cs.ucr.edu/~eamonn/SAX.htm) approach was

applied to discover the pattern similarity. SAS was customized with a window of 25

days, intersection of 20% and size of five alphabets. Then, by applying the distance

[23]

(3) min([(vec_cos(i, j)) for i, j in b])

a distance matrix was applied for all features to all polarities. Finally, the neutrals

were dropped and formed pair by pair by getting the maximum (max(a, b)) to every

cell within the matrix. Example of the resulted grouping is given in Fig. 5A, for spy

stories – Fig. 5B for War, and Fig. 5C ‒ for Criminal.

Fig. 5. Example of movies grouping based on the emotional fingerprints

4. Results and discussion

To evaluate the proposed approach, the IMDB dataset, which was used in phase three,

was split into a training set and a test set using the 10-fold cross validation[24]. Then,

the following standard metrics were applied: Precision, Recall, and F-Measure.

In more details, according to [25] Precision measures the ability of the

recommender system to retrieve as many relevant items as possible per request. It can

be calculated as follows:

(4)  





According to [26], a recall measures the system ability to retrieve fewer non-

relevant items as possible. It can be calculated as follows:

(5)  





According to [27], the F-Measures is a harmonic mean of both the precision and

recall. It can be calculated as follows:

(6)  





Additionally, the experiment was executed on four different user modes,

extracted from the social media at the first phase to understand how the recommender

is performing with different users’ moods. When applying the formulas (4), (5), and

(6) over the four different moods, the results were in Table 2.

Table 2. Evaluation metrics for the four moods

Metric/Mood

Anger

Joy

Sadness

Surprise

Precision

0.89

0.99

0.84

0.97

Recall

0.85

0.96

0.77

0.91

F-Measure (F1)

0.87

0.97

0.80

0.94

110

Table 2 shows the three metrics results with the different four moods. The results

show that the proposed model results varied according to the mood the user had. The

model achieved a higher precision with the following emotions in the following

sequence (Joy, Surprise, Anger, Sadness). Fig. 6 also depicts the different metrics

values for the different four moods as well.

Fig. 6. Average proposed system metrics

As it can be inferred from the figure above, the model achieved different results

according to the users’ moods. The proposed model achieved better precision results

with the Joy mood, then Surprise mood, Anger mood, and lastly Sadness mood.

To evaluate the performance of the proposed recommender, three baselines were

used to compare the proposed work with them. The first baseline is by comparing the

results of the proposed work with the results of the collaborative filtering technique

only. The second baseline is by comparing the proposed work results with the work

in [23] which depends on the EFBM only and ignores the mood and location

constraints. Finally, the third baseline is by comparing the proposed work results with

the work of [15]. In the last work, they applied the experiment where N=20.

Therefore, in order to be able to compare the proposed work with them, we applied

all the metrics where N=20 for all the baselines as well. After calculating the

aforementioned metrics on the proposed work at N=20, we have calculated the

average result for those metrics to be able to compare with the other baseline.

Table 3 shows the results of the N=20 of the four different moods before calculating

the average.

Table 3. Evaluation metrics for the four moods at N=20

Metric/Mood

Anger

Joy

Sadness

Surprise

Precision

0.91

0.97

0.8

0.94

Recall

0.86

0.91

0.78

0.89

F-Measure (F1)

0.88

0.94

0.79

0.91

Table 3 shows how the results of the three metrics applied at N=20 vary

according to the user’s mood as well. The model had a higher precision result with

0.2

0.4

0.6

0.8

1.2

Anger Joy Sadness Surprise

Average Metrics of the Proposed

System

Precision Recall F1

111

the Joy mood, then Surprise mood, Anger mood and finally the Sadness mood.

Fig. 7 also depicts the different metrics values for the different four moods as well.

Fig. 7. Evaluation metrics for the four moods at N=20

In addition, in work [15] they used the precision metric for evaluation only,

therefore we used this metric only in our comparison with the other baselines. Also,

they applied the precision metric in four different contexts. Therefore, we calculated

their precision average and we also calculated the precision average for our different

four moods where N=20. Table 4 shows the results of all the baselines where N=20.

Table 4. Comparison with Baselines at N=20

Method

Precision

Proposed Recommender

0.92

First Baseline

0.63

Second Baseline

0.87

Third Baseline

0.85

In Table 4, the precision value of the proposed recommender was 0.92. While

it was for the other three baselines as follows: the first baseline (Traditional CF)=63,

the second baseline (EFBM) = 0.87, while it was for the third baseline (work of

[15])=0.85. This reflects that the proposed recommender outperformed the other

baselines in the precision metric. Fig. 8 also shows a graphical representation of the

results.

Fig. 8. Precision metric comparison with baselines

0.2

0.4

0.6

0.8

1.2

Anger Joy Sadness Surprise

Metrics @N=20

Precision Recall F1

0.00

0.50

1.00

Proposed

Recommender

Frist Baseline Second

Baseline

Third Baseline

Precision Metric Comparison With Baselines

112

Fig. 8 depicts the results of comparing the results of the proposed recommender

with the other baselines. It reflects the superiority of the proposed system against the

other baselines.

The experimental results show that using a hybrid recommendation system that

takes in consideration the users’ moods, location, available time and the emotional

features in time, can improve the performance of the recommendation process, which

leads to better results and better user satisfaction. The comparison with the results of

other baselines shows that the proposed work outperformed the other baselines. This

is due to a better understanding of the user’s mood, the user’s availability, and

emotions, which in turn reflects the results of the recommendation, and the user’s

satisfaction. The average precision for the proposed recommender was 0.92, while it

achieved lower results in other baselines as follows: first baseline = 0.63, second

baseline = 0.87, and the third baseline = 0.85. For a deeper understanding of the

model, the proposed work was also experimentally tested with four different users’

moods, which was (Joy, Sadness, Surprise, and Anger). The experimental results

showed that the model best works with the following moods: (Joy, Surprise, Anger,

and Sadness) respectively. The results also, to some extent, reflect the nature of

human feelings. People can share and post about what they feel in their everyday life

especially when they are happy, as it was illustrated on the aforementioned example

of Table 1. Whereas when people feel sad, they are usually less expressive about their

feelings in social media. This can be due to psychological reasons or simply they are

not in a good mood to use the social media that some people still consider it as a

means of entertainment.

5. Conclusion and future work

The main objective of this paper is creating a hybrid recommender system that takes

into consideration many factors such as the user mood, the location, the user’s

availability, and the user’s emotions as well. The proposed recommender consists of

three phases. Each phase produces a list of movies that are filtered with other criteria

on the next phase. The first phase specifies the user’s mood, the next phase specifies

the spatio-temporal factors, and on the last phase, the emotional experience is

determined by applying the emotional fingerprint-based model.

Experimentations were done to understand the recommendation quality and

behavior. The results of the experiment were evaluated using three metrics of

evaluation: precision, recall, and F-Measure. In order to understand the quality of the

recommendation, the results were compared with three baselines. The first baseline

by comparing the results of the proposed model with the results of applying

collaborative filtering only. The second baseline is to compare the performance of

the upgraded system with the system without upgrade that presented in the previous

chapter of this thesis. The third baseline is to compare the experiment with another

hybrid recommender system. The comparison showed that the proposed work

outperformed all baselines, which reflects the importance of incorporating the user’s

mood, spatio-temporal aspects, and emotional fingerprint into the recommendation

system.

113

For future work, more user’s context should be taken in consideration such as

the weather, seasonality, user’s age, gender, job, educational level and more personal

details such as marital status, and even her parental status. Additionally, the model

should explore what happens if a user was impressed by a specific movie or a feature.

In other words, is it possible that the user can dislike a movie while he/she still loves

some features, which by the end lead him/her to watch that movie?

Further, adding the emotional factors opens many questions such as: is it still

possible that the user changes his/her mind to watch a different movie other than the

suggested by the recommender as a result of advertising? Is a feature preferred by a

user enough to let him/her watch that movie? Is there any effect for the mass

“euphoria” based on a specific topic to lead some people to watch a specific movie

or alternate their preferable movie to another one? Example, if some people prefer

the movie “Only Lovers Left Alive”, which has a strong “life and Death” topic, can

people look for some other “life and death” movie such as “Bicentennial Man”? Or

just to go by tags “movies by Jim Jarmusch” or “movie with Vampire” as it can be

seen on Fig. 9?

Fig. 9. IMDB proposal for some movies

114

R e f e r e n c e s

1. C h a m p i r i, Z. D., S. R. S h a h a m i r i, S. S. B. S a l i m. A Systematic Review of Scholar

Context-Aware Recommender Systems. – Expert Syst. Appl., Vol. 42, 2015, No 3,

pp. 1743-1758.

2. Shu, J., X. S h e n, H. Liu, B. Y i, Z. Z h a n g. A Content-Based Recommendation Algorithm for

Learning Resources. – Multimed. Syst., Vol. 24, 2018, No 2, pp. 163-173.

3. F u, M., H. Q u, D. M o g e s, L. L u. Attention Based Collaborative Filtering. – Neurocomputing,

Vol. 311, 2018, pp. 88-98.

4. W a n g, F., S. Lin, X. Luo, H. W u, R. W a n g, F. Z h o u. A Data-Driven Approach for Sketch-

Based 3D Shape Retrieval via Similar Drawing-Style Recommendation. – Comput. Graph.

Forum, Vol. 36, October 2017, No 7, pp. 157-166.

5. Chu, P. M., S. J. Lee. A Novel Recommender System for e-Commerce. – In: Proc. of 10th

International Congress on Image and Signal Processing, BioMedical Engineering and

Informatics (CISP-BMEI’17), 2017, pp. 1-5.

6. Z h a o, Q., C. W a n g, P. W a n g, M. Z h o u, C. J i a n g. A Novel Method on Information

Recommendation via Hybrid Similarity. – IEEE Trans. Syst. Man, Cybern. Syst., Vol. 48,

2018, No 3, pp. 448-459.

7. H a r i a d i, I., D. N u r j a n a h. Hybrid Attribute and Personality Based Recommender System for

Book Recommendation. – In: Proc. of International Conference on Data and Software

Engineering (ICoDSE’17), 2017, pp. 1-5.

8. Y a n g, S.-B., S.-H. S h i n, Y. J o u n, C. Koo. Exploring the Comparative Importance of Online

Hotel Reviews’ Heuristic Attributes in Review Helpfulness: A Conjoint Analysis Approach. –

J. Travel Tour. Mark., Vol. 34, September 2017, No 7, pp. 963-985.

9. W a n g, H., K. Guo. The Impact of Online Reviews on Exhibitor Behaviour: Evidence from Movie

Industry. – Enterp. Inf. Syst., Vol. 11, November 2017, No 10, pp. 1518-1534.

10. C l a r i z i a, F., F. C o l a c e, M. L o m b a r d i, F. P a s c a l e. A Context Aware Recommender

System for Digital Storytelling. – In: Proc. of IEEE 32nd International Conference on

Advanced Information Networking and Applications (AINA’18), 2018, pp. 542-549.

11. B o f f a, S., C. D. M a i o, B. G e r l a, M. P a r e n t e. Context-Aware Advertisment

Recommendation on Twitter through Rough Sets. – In: Proc. of IEEE International Conference

on Fuzzy Systems (FUZZ-IEEE’18), 2018, pp. 1-8.

12. Y a n g, Q. A Novel Recommendation System Based on Semantics and Context Awareness. –

Computing, Vol. 100, 2018, No 8, pp. 809-823.

13. D i x i t, V. S., P. J a i n. A Proposed Framework for Recommendations Aggregation in Context

Aware Recommender Systems. – In: Proc. of 8th International Conference on Cloud

Computing, Data Science & Engineering (Confluence), 2018, pp. 209-214.

14. C o l o m b o-M e n d o z a, L. O., R. V a l e n c i a-G a r c í a, A. R o d r í g u e z-G o n z á l e z,

G. A l o r-H e r n á n d e z, J. J. S a m p e r-Z a p a t e r. RecomMetz: A Context-Aware

Knowledge-Based Mobile Recommender System for Movie Showtimes. – Expert Syst. Appl.,

Vol. 42, 2015, No 3, pp. 1202-1222.

15. A b b a s, M., M. U. R i a z, A. R a u f, M. T. K h a n, S. K h a l i d. Context-Aware Youtube

Recommender System. – In: Proc. of International Conference on Information and

Communication Technologies (ICICT’17), 2017, pp. 161-164.

16. Cai, G., W. G u. Heterogeneous Context-Aware Recommendation Algorithm with Semi-

Supervised Tensor Factorization BT. – Intelligent Data Engineering and Automated Learning

– IDEAL’17, 2017, pp. 232-241.

17. Kim, D., C. P a r k, J. O h, S. Lee, H. Y u. Convolutional Matrix Factorization for Document

Context-Aware Recommendation. – In: Proc. of 10th ACM Conf. Recomm. Syst. (RecSys’16),

2016, pp. 233-240.

18. Y u, P., L. Lin, J. W a n g. A Novel Framework to Alleviate the Sparsity Problem in Context-Aware

Recommender Systems. – New Rev. Hypermedia Multimed., Vol. 23, April 2017, No 2,

pp. 141-158.

19. D e n g, S., D. W a n g, X. L i, G. X u. Exploring User Emotion in Microblogs for Music

Recommendation. – Expert Syst. Appl., Vol. 42, 2015, No 23, pp. 9284-9293.

115

20. Z h a o, Z., et al. Social-Aware Movie Recommendation via Multimodal Network Learning. – IEEE

Trans. Multimed., Vol. 20, 2018, No 2, pp. 430-440.

21. Z h a n g, Y., R. Jin, Z.-H. Z h o u. Understanding Bag-of-Words Model: A Statistical Framework.

– Int. J. Mach. Learn. Cybern., Vol. 1, 2010, No 1, pp. 43-52.

22. P. Ekman, R. J. Davidson, Eds. The Nature of Emotion: Fundamental Questions. New York, NY,

US, Oxford University Press, 1994.

23. N o s s h i, A., A. A s e m, M. B. S e n o u s y. Hybrid Recommender System Using Emotional

Fingerprints Model. – Int. J. Inf. Retr. Res., Vol. 9, 2019, No 3, p. Article 4 (in Press).

24. K o h a v i, R. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model

Selection. – In: Proc. of 14th International Joint Conference on Artificial Intelligence, Vol. 2,

1995, pp. 1137-1143.

25. Y u a n, T., H. W u, J. Zhu, L. S h e n, G. Q i a n. MS-UCF: A Reliable Recommendation Method

Based on Mood-Sensitivity Identification and User Credit. – In: Proc. of International

Conference on Information Management and Processing (ICIMP’18), 2018, pp. 16-20.

26. W a n g, H.-C., H.-T. J h o u, Y.-S. T s a i. Adapting Topic Map and Social Influence to the

Personalized Hybrid Recommender System. – Inf. Sci., NY, 2018.

27. Lim, H., H.-J. Kim. Item Recommendation Using Tag Emotion in Social Cataloging Services. –

Expert Syst. Appl., Vol. 89, 2017, pp. 179-187.

Received: 29.01.2019; Accepted: 14.02.2019 (fast track)