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

net/publication/329234471

Movie Recommendation System using Sentiment Analysis from

Microblogging Data

Preprint · November 2018

CITATIONS READS

0 3,080

4 authors, including:

Sudhanshu Kumar Kanjar De

7 PUBLICATIONS   160 CITATIONS   
Norwegian University of Science and Technology
24 PUBLICATIONS   207 CITATIONS   
SEE PROFILE
SEE PROFILE

Partha Roy
Indian Institute of Technology Roorkee
230 PUBLICATIONS   3,210 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Lexicon free approach for text recognition View project

A position and rotation invariant framework for sign language recognition (SLR) using Kinect View project

All content following this page was uploaded by Sudhanshu Kumar on 14 September 2020.

The user has requested enhancement of the downloaded file.

 Noname manuscript No.
(will be inserted by the editor)

Movie Recommendation System using Sentiment Analysis

from Microblogging Data

Sudhanshu Kumar · Shirsendu S. Halder ·

Kanjar De · Partha Pratim Roy

Received: date / Accepted: date

arXiv:1811.10804v1 [cs.IR] 27 Nov 2018

Abstract Recommendation systems are important intelligent systems that play a vi-
tal role in providing selective information to users. Traditional approaches in recom-
mendation systems include collaborative filtering and content-based filtering. How-
ever, these approaches have certain limitations like the necessity of prior user history
and habits for performing the task of recommendation. In order to reduce the effect
of such dependencies, this paper proposes a hybrid recommendation system which
combines the collaborative filtering, content-based filtering with sentiment analysis
of movie tweets. The movie tweets have been collected from microblogging web-
sites to understand the current trends and user response of the movie. Experiments
conducted on public database produce promising results.
Keywords Recommendation System · Twitter · Collaborative filtering · Content-
based filtering · Sentiment Analysis

1 Introduction

In today’s world, internet has become an important part of the human life. Users often
face the problem of excessive available information. Recommendation systems (RS)
are deployed to help users cope with this information explosion. RS are mostly used
in e-commerce applications and knowledge management systems such as tourism,
entertainment and online shopping portals. In this paper, we focus on RS for movies
are an important source of recreation and entertainment in our life. Movie suggestions
for users depend on web-based portals. Movies can be easily differentiated through
their genres like comedy, thriller, animation, and action. Another possible way to
categorize movies can be achieved on the basis of metadata such as year, language,
director or by cast. Most online video-streaming services provide a number of similar
Sudhanshu Kumar, Shirsendu S. Halder, Kanjar De, Partha Pratim Roy
Department of Computer Science and Engineering
Indian Institute of Technology, Roorkee, India.
E-mail: [email protected]
2 Sudhanshu Kumar et al.

movies to the user to utilising the user’s previously viewed or rated history. Movie
Recommendation Systems [23, 25, 48, 3, 53] help us to search our preferred movies
and also reduce the trouble of spending a lot of time searching for favorable movies.
The primary requirement of a movie recommendation system is that, it should be very
reliable and provide the user with the recommendation of movies which are similar
to their preferences. In recent times, with exponential increase in amount of online-
data, RS are very beneficial for taking decisions in different activities of day-to-day
life. RS are broadly classified into two categories: Collaborative filtering (CF) and
Content-based filtering (CBF).
It is a human tendency to take decisions on the basis of facts, predefined rules
and known information which is available over the internet and this inclination of
human behaviour gave rise to the concept of CF. Resnick et al. [37] introduced the
concept of CF in netnews, to help people find articles they liked in a huge stream
of available articles. CF help people make choices based on the perspective of other
people. Two users are considered like-minded when their rating for items are similar
whereas in CBF [46], items are suggested through the similarity among the contentual
information of the items. With the advent of numerous social media platforms like
Quora, Facebook, and Twitter, people are able to share their daily state of mind on the
internet. Twitter [1, 2, 13] is one of the most popular social media platform founded in
2006 where users can express their thoughts in limited characters. The Unique Selling
Proposition of Twitter is that the existing users not only receive information according
to their social links, but also gain access to other user-generated information. The
source of information on Twitter are called tweets which are of limited-character that
keep users updated about their favorite topics, people and movies. In this paper, we
propose a movie recommendation framework by fusing hybrid and sentiment scores
from MovieTweetings database.
The main contributions of the paper are as follows:
1. We propose a hybrid recommendation system by combining collaborative filter-
ing and content-based filtering.
2. Sentiment analysis is used to boost up this recommendation system.
3. A detailed analysis of proposed recommendation system is presented through
extensive experiment. Finally, a qualitative as well as quantitative comparison
with other baselines models is also demonstrated.
The rest of the paper is organized as follows: Section 2 summarized the related
work. The proposed methodology is presented in Section 3. Results obtained using
the proposed framework are shown in Section 4. Finally, the conclusion is drawn in
Section 5.

2 Related work

Recommender Systems are the most effective knowledge management systems that
help users to filter unusable data and deliver personalized ideas based on their past
historicial data and similar items which user are looking over the internet. Many RS
have been developed over the past decades. These systems used different approaches
Movie Recommendation System using Sentiment Analysis from Microblogging Data 3

like CF, CBF, hybrid and sentiment analysis system to recommend items. These are
discussed as follows.

2.1 Collaborative (CF), Content-based (CBF) and Hybrid filtering

Various RS approaches have been proposed in literature for recommending items

[41]. Primordial use of collaborative filtering was introduced in [16] which pro-
posed a search system based on document contents and responses collected from
other users. A RS for e-commerce applications by combining the pre-purchase and
post-purchase ratings was suggested by Guo et al. [18]. Soleymani et al. [43] worked
on an affective ranking of movie scenes, based on user’s emotion and video content
based features. They found that the movie scenes were correlated with the users self-
assessed arousal and valence. The peripheral physiological signals were utilized to
categorize and rank video contents.
Yang et al. [51] inferred implicit ratings from the number of pages the users read.
As many pages users read, the more the users are assumed to like the documents. This
concept is helpful to overcome the cold start problem in CF. Optimizing a RS is an
ill-posed problem. Researchers have proposed several optimization algorithms such
as gray wolf optimization [22], artificial bee colony [19], particle swarm optimiza-
tion [45] and genetic algorithms [6], etc. Katarya et al. [22] developed a movie rec-
ommendation system based on collaborative filtering which utilizes the bio-inspired
gray wolf optimizer and fuzzy c-mean (FCM) clustering techniques. The Gray wolf
optimizer was applied to obtain initial position of the cluster. The movie ratings were
further predicted based on user’s historical data and similarity of users. They im-
proved the existing framework in [21] proposing ABC-KM (artificial bee colony and
k-mean cluster) framework for collaborative movie recommendation system to re-
duce the scalability and cold start complication. This hybrid cluster and optimization
technique combination showed better accuracy in movie prediction compared to ex-
isting frameworks.
Content-based filtering [49, 27, 33, 47] is one of the most widely used and re-
searched recommendation system paradigm. This RS approach is based on the de-
scription of the item and a profile of the users preferences. In [31], the authors dis-
cussed about the discriminative power of words. They deduce that the title and ab-
stract are multiple times stronger than the body text and use a weighting scheme of
the title, abstract and the body text to retrieve relevant documents. Cantador et al.
[9] make use of user and item profiles, described in terms of weighted lists of social
tags to provide music recommendations. Meteran et al. [46] proposed a Personalized
Recommender System (PRES) to suggest articles for home improvement where the
similarity between the user profile vector and a document was determined by using
the combination of TF-IDF and the cosine similarity. Goossen et al. [17] proposed a
new method for recommending news items based on TF-IDF and a domain ontology
i.e. CF-IDF. The performance of this method outperformed the TF-IDF approach on
several measures such as accuracy, recall, and the F1 -measures when tested, evalu-
ated and implemented on Athena framework.
4 Sudhanshu Kumar et al.

Recent research has demonstrated that hybrid approach [34, 38, 44, 5, 7] is more
effective than traditional approaches. The main advantage of hybrid systems are that
they combine multiple recommendation techniques to mitigate the drawbacks of indi-
vidual techniques. In [30], the authors have developed a content-boosted CF System
which used pure content-based features in a collaborative framework. This system
further improved the prediction, first-rater and the sparsity problem. Zhang et al. [52]
developed a framework based on user recommender interaction that takes input from
the user, recommends N items to the user, and records user choice until none of the
recommended items favors. In [32], the authors developed a mobile recommender
system that combines a hybrid recommendation engine and a mobile 3D GIS archi-
tecture. For testing the proposed framework, twenty seven users were selected having
an age range of 24-48 years. To evaluate the performance of the system users were in-
structed to find restaurant, bars and accommodation while walking and driving along
a motorway. The user feedbacks demonstrated competent performance by the 3D
map-based interface that also overcomes the limited display size of most mobile de-
vices.

2.2 Sentiment analysis

Sentiment analysis [35, 8, 29, 36, 15, 50] is a widely used technique by researchers to
acquire people opinions. In [14], sentiment analysis has been used for rating products
for online software services. Their research enhances both CBF and CF algorithms,
using external reviews such as sentiment analysis and subjective logic. Sentiment
analysis technique has been used in [28] to calculate the polarity and confidence of
review sentences. The authors in [20] proposed Valence Aware Dictionary and Sen-
timent Reasoner (VADER) model for sentiment analysis. Lexical features were com-
bined for five general rules that embody grammatical and syntactical conventions for
expressing and emphasizing sentiment intensity. An F1 score of 0.96 was recorded
for classifying the tweets into positive, neutral, and negative classes. In [4], the author
proposed automatic feedback technique on the basis of data collected from Twitter.
Different classifiers such as Support Vector Machine, Naive Bayes, and Maximum
Entropy were used on twitter comments. The authors in [39] proposed music rec-
ommendation framework for mobile devices where recommendations of songs for
a user were based on the mood of user’s sentiment intensity. The studies were per-
formed on 200 participants (100 men and 100 women) to fill out their musical pref-
erence choice in his or her profile. Later, the participant’s profile was analyzed and
the results showed 91% user satisfaction rating. In [26], the author proposed KBridge
framework to solve the cold start problem in CF system. Sentiment analysis was also
used for microblogging posts in this framework and polarity score of the post was
assigned in 1 to 5 rating. The result showed an enhanced recommendation system by
bridging the gap between user communication knowledge between social networking
sites and program watching over TV. The author in [24] proposed a rating inference
approach to transform textual reviews into ratings to enable easy integration of senti-
ment analysis and CF.
Movie Recommendation System using Sentiment Analysis from Microblogging Data 5

Our proposed model is a hybrid recommender system whose results are boosted
using sentiment analysis score. Experimental evaluations, both quantitative and qual-
itative demonstrate the validity and effectiveness of our method.

3 Proposed system

The proposed sentiment based recommendation system is shown in Fig. 1. In this sec-
tion, we describe the different steps and components of the proposed recommender
system.

Fig. 1: Proposed movie recommendation framework.

3.1 Dataset description

The proposed system needs two different kinds of databases. One is a user-rated
movies database, where ratings for relevant movies are present and another is the
user-tweets from Twitter.

3.1.1 Public database

There are many popular public databases available which have been widely used in
the area of movie recommendation systems. In proposed framework, the tweets were
6 Sudhanshu Kumar et al.

extracted from Twitter for movies available in the database. In order to incorporate the
sentiment analysis in the proposed recommendation system, the tweets of relatively
new movies were used in this work.
Experiments conducted on public databases such as the Movielens 100K 1 , Movie-
lens 20M 2 , Internet Movie Database (IMDb3 ), and the Netflix database 4 were not
found suitable for our work. These databases were relatively outdated and contained
old movies whose appropriate microblogging data was not available. After thorough
assessment of different databases, the MovieTweetings database [12] was selected for
the proposed system.
MovieTweetings is widely considered as a modern version of MovieLens database.
The MovieTweetings database is an unfiltered database, unlike MovieLens where a
single user has rated at least 20 movies. The goal of this database is to provide an up-
to-date rating so it contains more realistic data for sentiment analysis. This database
is extracted from social media. It is very diverse but has a low sparsity value. Table 1
displays the relevant details of the MovieTweetings database.

Table 1: Details of MovieTweetings database.

Metric Value
Ratings 646410
Unique Users 51081
Unique Movies 29228
Start Year 1894
End Year 2017

3.1.2 Modified MovieTweetings database

In our proposed work, the MovieTweetings database is modified for implementing

the recommendation system. The main motivation for the modification was to use
sentiment analysis of tweets by users in the prediction of the recommendation of
movies. The MovieTweetings database contains movies ranging from years 1894 to
2017. Due to the scarcity of tweets for old movies, we only considered movies that
had a release year of more than 2014 and extracted a subset of the MovieTweetings
database complying with our motive.

release yearmovies ≥ 2014 (1)

After this database processing, a new tweaked database was formed for the im-
plementation of the recommendation system. This modified database consisted of
292863 ratings by 51081 users on 6209 different movies.
We find that 45% of the ratings consist of movies between 2014-2017 which is
around 20% of the total movies of the whole MovieTweetings database So in this
1 https://grouplens.org/datasets/movielens/100k/
2 https://grouplens.org/datasets/movielens/20m/
3 https://www.kaggle.com/orgesleka/imdbmovies
4 https://www.kaggle.com/netflix-inc/netflix-prize-data/data
Movie Recommendation System using Sentiment Analysis from Microblogging Data 7

work, only movies between 2014 -2017 has been used. The distribution of movies
with respect to the year of release is depicted in Fig. 2. The MovieTweetings database
has three different components. First component contains the mapping of users with
their Twitter IDs. The second component contains the ratings of movie items as rated
by the users and the final component contains information about the movies that were
rated.

Fig. 2: Year wise movies distribution.

The extracted database from MovieTweetings contains the ratings of movies by

users and their respective genres. However, it does not contain any other data except
for the release year and genres. Such data is proficient only in case of social filtering
where there are enough users with overlapping characteristics in the system. We can
use such data in case of collaborative filtering where suggestions are made solely
based on the ratings given by a pertaining user and items are recommended on the
judgement of similarity between users. In the proposed model, the socially filtered
data as well as the similarity of movies based on their attributes has been used. The
Movie Database (TMDb) API was used to get the attribute of the movies. TMDb 5 is a
premier source for an extensive metadata for movies that has more than 30 languages.
The movie attributes of MovieTweetings database are shown in Table 2.
The modified database contains very obscure movies from different countries and
different languages whose metadata was not available in TMDb. Such movies were
discarded that have almost nil metadata. The final database had around 4500 movies.
Table 3 shows the statistical details of the database.

5 https://www.themoviedb.org/about?language=en
8 Sudhanshu Kumar et al.

Table 2: Example of a movie entry in the modified MovieTweetings database.

Attribute Value
MovieID 0451279
Title Wonder Woman
Runtime 141 min
Genre Action,Adventure,Fantasy
Director Patty Jenkins
Writer Allan Heinberg
Actors Gal Gadot,Chris Pine
Rating 7.6 Massachusetts Institute of Technology in 1996.
Production Companies DC Films,Tencent Pictures
Popularity 524.772
Language en
Production Countries United States of America
Budget 816303142

Table 3: Details of the modified MovieTweetings database.

Metric Value
Ratings 292863
Unique Users 51081
Unique Movies 4515
Start Year 2014
End Year 2017

3.2 Analysis of user tweets

The tweets of movies from MovieTweetings were fetched through Twitter API 6 as
shown in Fig. 3. The extracted tweets consisted tremendous amounts of noise such as
hashtag, emoji, repetitive words and other irrelevant data which were removed using
preprocessing techniques.

3.2.1 Preprocessing of tweets

There were many noisy and uninformative parts in tweets, which did not contribute
much to accurate sentiment analysis. These include stop words, punctuations, web
links, special characters and repetitive words. Few example of noisy and uninfor-
mative parts are mentioned in Table 4. After preprocessing, the text extracted from
the tweets was used for sentiment analysis and then this information was used for
building the recommendation system.

Table 4: Examples of noisy and uninformative parts in tweets.

Types of noise Example
Stop words a, and, the, after, am
Lemma serve, served and serving
Web links www.tripadvisior.com
Filtering of repeating words happyyyy, heloooo
Special Characters !, @, #, $, %, and

6 https://developer.twitter.com/en/docs
Movie Recommendation System using Sentiment Analysis from Microblogging Data 9

3.2.2 Sentiment analysis of user tweets

VADER is a computationally efficient lexicon of sentiment-related words and rule-

based method that is used for sentiment analysis. In this approach, the detection of
subjectivity tweets can be handled by the sentiment process itself. VADER maps
words to sentiment by building a lexicon or a dictionary of sentiments. The dictio-
nary evaluates the sentiment of phrases and tweets. Lexical approaches can be em-
ployed to evaluate either the sentiment category or the score of the tweets. From each
tweet, VADER produces four sentiment components. Out of these the first three are:
positive, neutral, and negative. The fourth component is a compound score which is
a normalized score of first three metrics. The range of compound score varies from
-1 to 1, where -1 represents least preferred and 1 denotes most preferred choice of
movies. Later, the score is scaled in the range 1 − 10 using Eq. (2).

Fig. 3: A framework of the VADER sentiment analysis system.

Rating = [1 + (1 + x) × 2] × 2 (2)

where x is a compound score.

In addition, VADER executes faster in comparison to other methods such as Naive
Bayes analyzer, TextBlob. They work well in texts having slangs, emoticons, and
acronyms.
10 Sudhanshu Kumar et al.

3.3 Hybrid recommendation

In this section, we describe the combination of content-based similarity features

with collaborative social filtering to generate a hybrid recommendation model. Let
f = {f1 ,f2 ,· · · ,fn } and q = {q1 ,q2 ,· · · ,qn } are the content-based feature vectors and
weight vector, respectively.
We construct the closeness C of two items i and j as:
(P
N
n=1 fn (Ani , Anj ), for i 6= j
C(i, j) = (3)
0, otherwise
where fn (Ani , Anj ), corresponds to the similarity between feature values Ani
and Anj corresponding to two items. In Eq. (3) the closeness of the items is deter-
mined using the metadata or the relevant information related to the items. Fij are
constructed by combining the closeness vector C for all the items and multiplying it
with the weight vectors q. Fij is a feature matrix of dimension n × M (M2 −1) , where
n and M are the number of feature attributes and number of items, respectively.
The weight vectors q are evaluated using a social graph of items that indicate the
user likeness of items. Let U = {u1 , u2 , · · · , un } where ui is a user in the database.
A User-Item matrix is constructed for M items. An important property of the User-
Item matrix is that it has very high sparsity. Typical collaborative Filtering [42] uses
this User-Item matrix to predict a user’s rating of a particular item i by analyzing
the ratings of other users in the user’s neighborhood, normally a K neighboring users.
Neighboring users are recognized by similarity measures like Cosine Similarity, Pear-
son Correlation etc. After selecting K neighbouring users, the weighted aggregation
of the ratings are as follow:
1 X
rating(u, i) = similarity(useru , uservk ) · ratingvki (4)
K
k∈K

where u and vk are target user and K nearest neighbors, respectively. The pro-
cedure of collaborative filtering is used to overcome the sparsity of the User-Item
matrix instead of directly using it to predict ratings. We employ the tweaked User-
Item matrix to construct a social graph using items as nodes. This graph represents
user perception of similarity between items. The determination of feature weights
complies with the social graph.
In order to determine the optimal feature weights q, we formulate a framework as
described in Eq. (5):
S(i, j) = q · Fij (5)
which can be expanded as:

S(i, j) = q1 · f1 (A1i , A1j ) + q2 · f2 (A2i , A2j ) + · · · + qn · fn (Ani , Anj ) (6)

Now for two items i and j, S(i, j) is evaluated from the User-Item matrix, where
S(i, j) are the number of users who are interested in both items i, j . For the entire
database, S is a matrix of dimension 1 × M (M2 −1) and q is a matrix of dimension
1 × n, where n is the number of content-based features and dimenstionality of F is
Movie Recommendation System using Sentiment Analysis from Microblogging Data 11

n × M (M2 −1) . We calculate the weight vector q for all the metadata feature attributes
for the complete items using the Moore-Penrose Pseudoinverse as in Eq. (7) :

q = S −1 · F (7)

3.4 Weighted score fusion

In section 3.3, we derived the weights q of the feature vectors using both movie-
similarity and user-similarity paradigms. These weights q are normalized between [0,
1] and the concept of sentiment-fusion is utilized in the proposed system. Through
the retrieved user tweets, a sentiment rating is fabricated for all movies M . Let S ∈
{s1 , s2 , · · · , sn } where si is the rating of movie i calculated using Eq. (2). A function
G(i, j) for two movies i, j are defined based on their sentiment ratings si and sj as
mentioned in Eq. (8):

G(i, j) = D − |si − sj | (8)

where D is a constant. The constant D in Eq. (8) is taken as 10 because the ratings
are in a scale of 1 − 10. Another function H(i, j) defined as:

H(i, j) = q · fij (9)

where fij is the feature similarity between movies i, j and q are the set of optimal
weights as determined by Eq. (7). The final combined similarity CS(i, j) is described
in Eq. (10). It is a weighted combination of the defined functions G and H.

CS(i, j) = ω1 · H(i, j) + ω2 · G(i, j) (10)

ω1 + ω2 = 1, ω1 , ω2 ∈ [0, 1] (11)

where ω1 corresponds to the weight of the similarity score calculated from the
hybrid model and ω2 corresponds to the weight of the sentiment similarity score.

4 Experimental results and analysis

In this section, the different correlation coefficient results are presented between sen-
timent ratings and IMDb movie ratings. We also describe a comprehensive quantita-
tive and qualitative analysis, illustrating the efficacy and the precision of our proposed
framework on movie database.
12 Sudhanshu Kumar et al.

4.1 Correlation between sentiment and IMDb movie ratings

We conducted the statistical analysis between sentiment ratings X and movie rating
Y to find the correlation coefficient. The correlation coefficient value varies from
-1 to +1. Let D denote a database of movies and N denotes the number of total
movies in the database. The statistical correlation coefficients are as follows: Spear-
man Rank Order, Correlation Coefficient (SROCC), Kendall Rank Correlation Co-
efficient (KRCC) and Pearson Linear Correlation Coefficient (P LCC). Table 5 dis-
plays the values of different co-relation coefficients utilized by us. In our experiments,
we have found that sentiment and movie ratings are positively correlated. For PLCC,
xi and yi are sentiment rating and IMDb movie rating, respectively for ith movie
while x̄ denotes the mean sentiment score and ȳ denotes the mean movie rating in the
database. For SROCC, di is the difference between the sentiment rating and movie
rating of the ith movie in the database. For KRCC, Nc and Nd represents the number
of concordant and discordant pairs in the database, respectively.

Table 5: Correlation measures between sentiment and movie ratings.

Correlation coefficient Definition Value

PN
i=1 (xi −x̄)(y i −ȳ)
PLCC qP q 0.76
N (x −x̄)2 PN (y −ȳ)2
i=1 i i=1 i
N
6
d2i
P
SROCC 1− N (N 2 −1)
0.72
i=1
2(Nc −Nd )
KRCC N (N −1)
0.51

4.2 Evaluation metric

In many real-world applications, relevant recommendations are suggested by the sys-

tem, instead of directly predicting rating values. This is known as Top-N recommen-
dations [11, 10, 40] and suggests specific items to users that are likeable. Normally,
in majority of the common literature RMSE (Root Mean Squared Error) or MSE
(Mean Squared Error) are used. But such error metrics are less significant for Top-N
recommendation system. Therefore, the direct alternative methodologies are used for
evaluation metric (e.g. precision). Precision is defined in terms of movies that are
relevant and the movies that are recommended by the model. Let Lrel be the relevant
movies and Lrec be the recommended movies. In proposed system, Precision@N is
defined as follow in Eq. (12):

Lrel ∩ Lrec
P recision@N = (12)
Lrec
For the proposed model, the choice of weights in the fusion in Eq. (10) is deter-
mined by evaluating the Precision@5 and Precision@10 for a different combination
of weights ω1 and ω2 conforming with Eq. (11).
Movie Recommendation System using Sentiment Analysis from Microblogging Data 13

4.3 Weight selection for weighted fusion

For every movie, Top-N recommendation list is evaluated using Eq. (10). The choice
of the weights ω1 and ω2 in Eq. (10) are decided by experiments conducted on the
metric mentioned in the section 4.2. The Precision@N is evaluated as in Eq. (12). The
recommendations of all movies are collected from public databases like IMDb and
TMDb. The recommended movies by these two sources are consider as the ground-
truth. We compare the results of the Precision@5 and Precision@10 for different
values of ω1 and ω2 . We choose the values of ω1 and ω2 for which the precision
values are the best.

Fig. 4: Precision of Top-5 and Top-10 movies with varying sentiment similarity
weights.

From Fig. 4, an observation can be made that the maximum precision for weight
values are between 0.5 and 0.6. Hence, ω1 and ω2 values are selected 0.5 in the
proposed system.

4.4 Comparative analysis

In this section, we present a comparative analysis of our proposed system with base-
line models. The base-line models used for the purpose of our experiments are Pure
Hybrid Model and Sentiment Similarity model. The Pure Hybrid Model (PH Model)
is a combination of content-based filtering and collaborative filtering where both
these filtering methods supplement each other to increase the robustness of the com-
plete model. The recommended movies are based on the similarity of attributes like
the genre, director, actor, etc. The similarities are evaluated using weights obtained
using a social graph as described in Section 3.3. Sentiment Similarity model (SS
Model) recommends the movie based solely on the similarity of the movie tweets
14 Sudhanshu Kumar et al.

of the corresponding tuple of movies. We evaluate our proposed method using Pre-
cision@5 and Precision@10. Fig. 5 displays the quantitative comparative results of
our proposed system with the baseline models. For Precision@5, the average preci-
sion value of SS Model and PH Model are 0.54 and 1.86, respectively. Similarly, for
Precision@10, the average precision value of SS Models and PH Model are 1.04 and
3.31, respectively. Our proposed model achieves a better precision value in both cases
with 2.54 for Top-5 and 4.97 for Top-10 in comparison to PH and SS Model. Thus we
can infer that our method will suggest at least 2 recommended movies out of 5 and
5 recommended movies out of 10 considering the general consensus as validated by
IMDb and TMDb databases.

Fig. 5: Comparison of proposed model with baselines models.

4.5 Qualitative analysis

In this section, we show the qualitative results of some of the recommended movies
given by the proposed system. It also includes movies from other languages as shown
in the Tables 6 and 7. It is clear from these tables that the recommendations from
the proposed system have many intersecting movies both with the recommendations
from IMDb and TMDb. The qualitative analysis of Hollywood as well as Bollywood
(Indian) movies show the capability of suggesting recommendations in diverse situa-
tions.

5 Conclusion and future works

Recommender systems are an important medium of information filtering system in

the modern age where there are enormous amounts of data readily available. In this
Movie Recommendation System using Sentiment Analysis from Microblogging Data 15

Table 6: Qualitative analysis of Wonder Woman movie: Language (English). Movies

in bold are intersecting with either IMDb or TMDb
Recommendations from the
IMDb TMDb
proposed system
Justice League Guardians of the Galaxy Vol. 2 Batman v Superman: Dawn of Justice
Batman v Superman: Dawn of Justice Spider-Man: Homecoming Suicide Squad
Suicide Squad Logan Thor: Ragnarok
Thor: Ragnarok Thor: Ragnarok Justice League
Spider-Man: Homecoming Justice League Warcraft
Pirates of the Caribbean:
Deadpool Doctor Strange
Dead Men Tell No Tales
Logan Doctor Strange Guardians of the Galaxy Vol. 2
Captain America: Civil War Baby Driver Kong: Skull Island
Doctor Strange Kong: Skull Island The LEGO Batman Movie
Guardians of the Galaxy Vol. 2 Life Batman and Harley Quinn

Table 7: Qualitative analysis of Neerja movie: Language (Hindi). Movies in bold are
intersecting with either IMDb or TMDb
IMDb TMDb Recommendations from the proposed system
Airlift Airlift Simran
Pink Pink Fan
Kapoor & Sons Rustom Raabta
Udta Punjab Ghayal Once Again Udta Punjab
Drishyam Mary Kom Rocky Handsome
Rustom Udta Punjab Rangoon
M.S. Dhoni: The Untold Story Force 2 Raabta
Raabta Fan Force 2
Dear Zindagi Rocky Handsome Te3n
Rangoon Simran Airlift

paper, we have proposed a recommender system which uses sentiment analysis data
from twitter along with movie metadata and a social graph to recommend movies.
Sentiment analysis provide information about how audience is reacting to a particu-
lar movie and this information has been observed to be useful. The proposed system
used weighted score fusion to improve the recommendations. Based on our exper-
iments, the average precision in Top-5 and Top-10 for sentiment similarity, hybrid,
and proposed model are .54 & 1.04, 1.86 & 3.31, and 2.54 & 4.97, respectively. We
found that proposed model recommends more precisely than the others models. In
future, we look forward to extract more information about emotional tone of the user
from different social media platforms to further improve the recommendation sys-
tem using sentiment analysis. The findings of this paper show that there is enough
scope for research to explore emotion information about users and integrate them in
recommendation system for e-commerce application. Another direction of research
will be to update weights derived from the social graph in real-time. We have used
a static database for our experiments where we only considered movies up to 2017.
This framework can be explored in a dynamic paradigm where the inclusion of recent
movies could be made regular.
16 Sudhanshu Kumar et al.

Compliance with ethical standards

Conflict of interest The authors declared that they have no conflicts of interest to this
work.

References

1. Fabian Abel, Qi Gao, Geert-Jan Houben, and Ke Tao. Analyzing user modeling
on twitter for personalized news recommendations. In International Conference
on User Modeling, Adaptation, and Personalization, pages 1–12. Springer, 2011.
2. Fabian Abel, Qi Gao, Geert-Jan Houben, and Ke Tao. Twitter-based user mod-
eling for news recommendations. In International Joint Conference on Artificial
Intelligence, volume 13, pages 2962–2966, 2013.
3. Gediminas Adomavicius and Alexander Tuzhilin. Toward the next generation of
recommender systems: A survey of the state-of-the-art and possible extensions.
IEEE Transactions on Knowledge and Data Engineering, 17(6):734–749, 2005.
4. Minara P Anto, Mejo Antony, KM Muhsina, Nivya Johny, Vinay James, and
Aswathy Wilson. Product rating using sentiment analysis. In International Con-
ference on Electrical, Electronics, and Optimization Techniques, pages 3458–
3462. IEEE, 2016.
5. Ehsan Aslanian, Mohammadreza Radmanesh, and Mahdi Jalili. Hybrid recom-
mender systems based on content feature relationship. IEEE Transactions on
Industrial Informatics, 2016.
6. Jesus Bobadilla, Fernando Ortega, Antonio Hernando, and Javier Alcalá. Im-
proving collaborative filtering recommender system results and performance us-
ing genetic algorithms. Knowledge-based systems, 24(8):1310–1316, 2011.
7. Robin Burke. Hybrid recommender systems: Survey and experiments. User
modeling and user-adapted interaction, 12(4):331–370, 2002.
8. Erik Cambria. Affective computing and sentiment analysis. IEEE Intelligent
Systems, 31(2):102–107, 2016.
9. Iván Cantador, Alejandro Bellogı́n, and David Vallet. Content-based recommen-
dation in social tagging systems. In Proceedings of the Fourth Conference on
Recommender systems, pages 237–240. ACM, 2010.
10. Paolo Cremonesi, Yehuda Koren, and Roberto Turrin. Performance of recom-
mender algorithms on top-n recommendation tasks. In Proceedings of the Fourth
Conference on Recommender Systems, pages 39–46. ACM, 2010. ISBN 978-1-
60558-906-0.
11. Mukund Deshpande and George Karypis. Item-based top-n recommendation
algorithms. ACM Transactions on Information Systems, 22(1):143–177, 2004.
ISSN 1046-8188.
12. Simon Dooms, Toon De Pessemier, and Luc Martens. Movietweetings: a movie
rating dataset collected from twitter. In Workshop on Crowdsourcing and Human
Computation for Recommender Systems, 2013.
13. Xin Fu and Yun Shen. Study of collective user behaviour in twitter: a fuzzy
approach. Neural Computing and Applications, 25(7-8):1603–1614, 2014.
Movie Recommendation System using Sentiment Analysis from Microblogging Data 17

14. Lahiru S Gallege and Rajeev R Raje. Towards selecting and recommending
online software services by evaluating external attributes. In Eleventh Annual
Cyber and Information Security Research Conference, page 23, 2016.
15. Himaanshu Gauba, Pradeep Kumar, Partha Pratim Roy, Priyanka Singh,
Debi Prosad Dogra, and Balasubramanian Raman. Prediction of advertisement
preference by fusing eeg response and sentiment analysis. Neural Networks, 92:
77–88, 2017.
16. David Goldberg, David Nichols, Brian M Oki, and Douglas Terry. Using collab-
orative filtering to weave an information tapestry. Communications of the ACM,
35(12):61–70, 1992.
17. Frank Goossen, Wouter IJntema, Flavius Frasincar, Frederik Hogenboom, and
Uzay Kaymak. News personalization using the cf-idf semantic recommender.
In Proceedings of the International Conference on Web Intelligence, Mining and
Semantics, page 10. ACM, 2011.
18. Guibing Guo and Mohamed Elgendi. A new recommender system for 3d e-
commerce: An EEG based approach. Journal of Advanced Management Science,
1(1):61–65, 2013.
19. Chia-Cheng Hsu, Hsin-Chin Chen, Kuo-Kuang Huang, and Yueh-Min Huang. A
personalized auxiliary material recommendation system based on learning style
on facebook applying an artificial bee colony algorithm. Computers & Mathe-
matics with Applications, 64(5):1506–1513, 2012.
20. Clayton J Hutto and Eric Gilbert. Vader: A parsimonious rule-based model for
sentiment analysis of social media text. In Eighth International Conference on
Weblogs and Social Media, 2014.
21. Rahul Katarya. Movie recommender system with metaheuristic artificial bee.
Neural Computing and Applications, pages 1–8, 2018.
22. Rahul Katarya and Om Prakash Verma. Recommender system with grey wolf
optimizer and fcm. Neural Computing and Applications, pages 1–9, 2016.
23. Rahul Katarya and Om Prakash Verma. An effective collaborative movie rec-
ommender system with cuckoo search. Egyptian Informatics Journal, 18(2):
105–112, 2017.
24. Cane WK Leung, Stephen CF Chan, and Fu-lai Chung. Integrating collaborative
filtering and sentiment analysis: A rating inference approach. In Proceedings of
the ECAI workshop on recommender systems, pages 62–66, 2006.
25. Fangfang Li, Guandong Xu, and Longbing Cao. Two-level matrix factorization
for recommender systems. Neural Computing and Applications, 27(8):2267–
2278, 2016.
26. Hui Li, Jiangtao Cui, Bingqing Shen, and Jianfeng Ma. An intelligent movie
recommendation system through group-level sentiment analysis in microblogs.
Neurocomputing, 210:164–173, 2016.
27. Pasquale Lops, Marco De Gemmis, and Giovanni Semeraro. Content-based rec-
ommender systems: State of the art and trends. In Recommender Systems Hand-
book, pages 73–105. Springer, 2011.
28. Steven Loria. Textblob: simplified text processing. Secondary TextBlob: Simpli-
fied Text Processing, 2014.
18 Sudhanshu Kumar et al.

29. Walaa Medhat, Ahmed Hassan, and Hoda Korashy. Sentiment analysis algo-
rithms and applications: A survey. Ain Shams Engineering Journal, 5(4):1093–
1113, 2014.
30. Prem Melville, Raymond J Mooney, and Ramadass Nagarajan. Content-boosted
collaborative filtering for improved recommendations. Aaai/iaai, 23:187–192,
2002.
31. Cristiano Nascimento, Alberto HF Laender, Altigran S da Silva, and Mar-
cos André Gonçalves. A source independent framework for research paper rec-
ommendation. In Proceedings of the Eleventh Annual International Joint Con-
ference on Digital libraries, pages 297–306. ACM, 2011.
32. José M Noguera, Manuel J Barranco, Rafael J Segura, and Luis Martı́Nez. A
mobile 3d-gis hybrid recommender system for tourism. Information Sciences,
215:37–52, 2012.
33. Simon Philip, P Shola, and A Ovye. Application of content-based approach
in research paper recommendation system for a digital library. International
Journal of Advanced Computer Science and Applications, 5(10), 2014.
34. Carlos Porcel, A Tejeda-Lorente, MA Martı́nez, and Enrique Herrera-Viedma. A
hybrid recommender system for the selective dissemination of research resources
in a technology transfer office. Information Sciences, 184(1):1–19, 2012.
35. Rudy Prabowo and Mike Thelwall. Sentiment analysis: A combined approach.
Journal of Informetrics, 3(2):143–157, 2009.
36. Kumar Ravi and Vadlamani Ravi. A survey on opinion mining and sentiment
analysis: tasks, approaches and applications. Knowledge-Based Systems, 89:14–
46, 2015.
37. Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John
Riedl. Grouplens: an open architecture for collaborative filtering of netnews. In
Proceedings of the Conference on Computer Supported Cooperative Work, pages
175–186. ACM, 1994.
38. Ade Romadhony, Said Al Faraby, and Bambang Pudjoatmodjo. Online shop-
ping recommender system using hybrid method. In International Conference of
Information and Communication Technology, pages 166–169. IEEE, 2013.
39. Renata L Rosa, Demsteneso Z Rodriguez, and Graça Bressan. Music recommen-
dation system based on user’s sentiments extracted from social networks. IEEE
Transactions on Consumer Electronics, 61(3):359–367, 2015.
40. Alan Said, Alejandro Bellogı́n, and Arjen De Vries. A top-n recommender sys-
tem evaluation protocol inspired by deployed systems. In LSRS Workshop at
ACM RecSys. Citeseer, 2013.
41. Vivek Sembium, Rajeev Rastogi, Atul Saroop, and Srujana Merugu. Recom-
mending product sizes to customers. In Proceedings of the Eleventh Conference
on Recommender Systems, pages 243–250. ACM, 2017.
42. Yue Shi, Martha Larson, and Alan Hanjalic. Collaborative filtering beyond the
user-item matrix: A survey of the state of the art and future challenges. ACM
Computing Surveys, 47(1), 2014.
43. Mohammad Soleymani, Guillaume Chanel, Joep JM Kierkels, and Thierry Pun.
Affective ranking of movie scenes using physiological signals and content anal-
ysis. In Second Workshop on Multimedia Semantics, pages 32–39, 2008.
 Movie Recommendation System using Sentiment Analysis from Microblogging Data 19

44. Stephan Spiegel. A hybrid approach to recommender systems based on ma-

trix factorization. Department for Agent Technologies and Telecommunications,
Technical University Berlin, 2009.
45. Supiya Ujjin and Peter J Bentley. Particle swarm optimization recommender sys-
tem. In Proceedings of Swarm Intelligence Symposium, pages 124–131. IEEE,
2003.
46. Robin Van Meteren and Maarten Van Someren. Using content-based filtering for
recommendation. In Proceedings of the Machine Learning in the New Informa-
tion Age: Workshop, pages 47–56, 2000.
47. Tiphaine Viard and Raphaël Fournier-S’niehotta. Movie rating prediction using
content-based and link stream features. arXiv preprint arXiv:1805.02893, 2018.
48. Xin Wan and Toshio Okamoto. Utilizing learning process to improve recom-
mender system for group learning support. Neural Computing and Applications,
20(5):611–621, 2011.
49. Mohammed Wasid and Vibhor Kant. A particle swarm approach to collaborative
filtering based recommender systems through fuzzy features. Procedia Computer
Science, 54:440–448, 2015.
50. Mahendra Yadava, Pradeep Kumar, Rajkumar Saini, Partha Pratim Roy, and
Debi Prosad Dogra. Analysis of eeg signals and its application to neuromar-
keting. Multimedia Tools and Applications, 76(18):19087–19111, 2017.
51. Chenxing Yang, Baogang Wei, Jiangqin Wu, Yin Zhang, and Liang Zhang.
Cares: a ranking-oriented cadal recommender system. In Proceedings of the
Ninth Joint Conference on Digital libraries, pages 203–212. ACM, 2009.
52. Heng-Ru Zhang, Fan Min, Xu He, and Yuan-Yuan Xu. A hybrid recommender
system based on user-recommender interaction. Mathematical Problems in En-
gineering, 2015.
53. Yao Zhou, Lei Zhang, and Zhang Yi. Predicting movie box-office revenues using
deep neural networks. Neural Computing and Applications, pages 1–11, 2017.

View publication stats