Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Movie Recommendation System using Content-

based Filtering
Yeole Madhavi B.1, Rokade Monika D.2, Khatal Sunil S.3
1
Student, Department of Computer Engineering, SPCOE Dumbarwadi (Otur), Maharashtra, India
2&3
Assistant Professor, Department of Computer Engineering, SPCOE Dumbarwadi (Otur),
Maharashtra, India

ABSTRACT
There is already enough content available on the movie recommendation system. Showing the movie
recommendations is essential so that the user need not waste a lot of time searching for the content which he/she
might like. Thus, movie recommendation system plays a vital role to get user personalized movie recommendations.
After searching a lot on the internet and referring to a lot of research papers, we got to know that the
recommendations made using Content-based Filtering are using a single text to vector conversion technique and a
single technique to find the similarity between the vectors. In this research work, we have used multiple text to
vector conversion techniques and manipulated the results of the multiple algorithms to get the final recommendation
list. You can think of it as a hybrid approach using the Content-based Filtering technique only.

Keyword: - Movie Recommendations, Content-based Filtering, Text to vector, Vector similarity, Hybrid
approach

1. INTRODUCTION

Due to abundance of information collected till 21st century and the increasing rate of information flowing over the
internet, there is a lot of confusion related to what to consume and what not to consume. Even on YouTube, when
you want to watch a video of a particular concept, generally, there are a lot of videos available out there for you.
Now, since the results are ranked appropriately, there may not be much issue but what if the results were not ranked
appropriately? Well, in that case, we would probably spend a lot of time to find the best possible video which suits
us and satisfies our need. This recommendation results are when you search something on a website. Next time,
when you visit a particular website, without even searching, sometimes the system is able to show you
recommendations which you might like. Isn’t this an interesting feature? So, basically, the job of a recommender
system is to suggest the most relevant items to the user. Recommendation systems are used in YouTube for video
recommendation, Amazon and Flipkart for product recommendation, Netflix and Amazon Prime for movie
recommendation, and so on. Whatever you do on such websites, there is a system which see your behavior and then
ultimately suggest things / items with which you are highly likely to engage. This research paper deals with movie
recommendations and logic behind movie recommendation system, traditional movie recommendation systems,
issues related to traditional movie recommendation systems, and a proposed solution for Artificial Intelligence based
personalized movie recommendation system. A lot of famous movie recommendation related datasets are already
available on Kaggle and other websites. Some of the famous datasets include Movielens dataset, TMDB Movie
Dataset, and the dataset by Netflix itself. Websites like Netflix, Amazon Prime, etc. use movie recommendation to
increase their revenue or profits by ultimately improving the user experience. In fact, there was a competition
conducted by Netflix in the year 2009 with a prize money of nearly 1 million dollars ($1M) for making at least 10%
improvement in the existing system.
As dealt earlier, we have a lot of data available at our exposure and we need to filter the data in order to consume it
because generally we are not interested in each and everything available to us. In order to filter the data, we need
some filtering techniques. There are different types of filtering techniques or movie recommendation algorithms
over which a recommendation system can be based upon.
Major filtering techniques or movie recommendation algorithms are as follows:

14954 www.ijariie.com 633

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

1. Content Based Filtering

2. Collaborative Filtering
3. Hybrid Filtering
Some of these techniques can be further broken into subparts

2. LITERATURE REVIEW
Sang-Min Choi, et. al. [1] mentioned about the shortcomings of collaborative filtering approach like sparsity
problem or the cold-start problem. In order to avoid this issue, the authors have proposed a solution to use category
information. The authors have proposed a movie recommendation system which is based on genre correlations. The
authors stated that the category information is present for the newly created content. Thus, even if the new content
does not have enough ratings or enough views, still it can pop up in the recommendations list with the help of
category or genre information. The proposed solution is unbiased over the highly rated most watched content and
new content which is not watched a lot. Hence, even a new movie can be recommended by the recommendation
system.
George Lekakos, et. al. [2] proposed a solution of movie recommendation using hybrid approach. The authors stated
that Content based filtering and Collaborative filtering have their own shortcomings are can be used in a specific
situation. Hence, the authors have come up with a hybrid approach which takes into consideration both content-
based filtering as well as collaborative filtering. The solution is implemented in 'MoRe' which is a movie
recommendation system. For the sake of pure collaborative filtering, Pearson correlation coefficient has not been
used. Instead, a new formula has been used. But this formula has an issue of 'divide by zero' error. This error occurs
when the users have given same rating to the movies. Hence, the authors have ignored such users. In case of pure
content-based recommendation system, the authors have used cosine similarity by taking into consideration movie
writers, cast, directors, producers and the movie genre. The authors have implemented a hybrid recommendation
method by using 2 variations - 'substitute' and 'switching'. Both of these approaches show results based on
collaborative filtering and show recommendations based on content-based filtering when a certain criterion is met.
Hence, the authors use collaborative filtering technique as their main approach.
Debashis Das, et. al. [3] wrote about the different types of recommendation systems and their general information.
This was a survey paper on recommendation systems. The authors mentioned about Personalized recommendation
systems as well as non-personalized systems. User based collaborative filtering and item based collaborative
filtering was explained with a very good example. The authors have also mentioned about the merits and demerits of
different recommendation systems.
Jiang Zhang, et. al. [4] proposed a collaborative filtering approach for movie recommendation and they named their
approach as 'Weighted KM-Slope-VU'. The authors divided the users into clusters of similar users with the help of
K-means clustering. Later, they selected a virtual opinion leader from each cluster which represents the all the users
in that particular cluster. Now, instead of processing complete user-item rating matrix, the authors processed virtual
opinion leader-item matrix which is of small size. Later, this smaller matrix is processed by the unique algorithm
proposed by the authors. This way, the time taken to get recommendations is reduced.
S. Rajarajeswari, et. al. [5] discussed about Simple Recommender System, Content-based Recommender System,
Collaborative Filtering based Recommender System and finally proposed a solution consisting of Hybrid
Recommendation System. The authors have taken into consideration cosine similarity and SVD. Their system gets
30 movie recommendations using cosine similarity. Later, they filter these movies based on SVD and user ratings.
The system takes into consideration only the recent movie which the user has watched because the authors have
proposed a solution which takes as input only one movie.
Muyeed Ahmed, et. al. [6] proposed a solution using K-means clustering algorithm. Authors have separated similar
users by using clusters. Later, the authors have created a neural network for each cluster for recommendation
purpose. The proposed system consists of steps like Data Preprocessing, Principal Component Analysis, Clustering,
Data Preprocessing for Neural Network, and Building Neural Network. User rating, user preference, and user
consumption ratio have been taken into consideration. After clustering phase, for the purpose of predicting the
ratings which the user might give to the unwatched movies, the authors have used neural network. Finally,
recommendations are made with the help of predicted high ratings.
Gaurav Arora, et. al. [7] have proposed a solution of movie recommendation which is based on users' similarity. The
research paper is very general in the sense that the authors have not mentioned the internal working details. In the
Methodology section, the authors have mentioned about City Block Distance and Euclidean Distance but have not
mentioned anything about cosine similarity or other techniques. The authors stated that the recommendation system

14954 www.ijariie.com 634

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

is based on hybrid approach using context based filtering and collaborative filtering but neither they have stated
about the parameters used, not they have stated about the internal working details.
V. Subramaniyaswamy, et. al. [8] have proposed a solution of personalized movie recommendation which uses
collaborative filtering technique. Euclidean distance metric has been used in oder to find out the most similar user.
The user with least value of Euclidean distance is found. Finally, movie recommendation is based on what that
particular user has best rated. The authors have even claimed that the recommendations are varied as per the time so
that the system performs better with the changing taste of the user with time.
Harper, et. al. [9] mentioned the details about the MovieLens Dataset in their research paper. This dataset is widely
used especially for movie recommendation purpose. There are different versions of dataset available like MovieLens
100K / 1M / 10M / 20M / 25M /1B Dataset. The dataset consists of features like user id, item id / movie id, rating,
timestamp, movie title, IMDb URL, release date, etc. along with the movie genre information.
According to R. Lavanya, et. al. [10], in order to tackle the information explosion problem, recommendation
systems are helpful. Authors mentioned about the problems of data sparsity, cold start problem, scalability, etc.
Authors have done a literature review of nearly 15 research papers related to movie recommendation system. After
reviewing all these papers, they observed that most of the authors have used collaborative filtering rather than
content-based filtering. Also, the authors noticed that a lot of authors have used hybrid-based approach. Even though
a lot of research has been done on recommendation systems, there is always a scope for doing more in order to solve
the existing drawbacks.
Ms. Neeharika Immaneni, et. al. [11] proposed a hybrid recommendation technique which takes into consideration
both content-based filtering approach as well as collaborative filtering approach in a hierarchial manner in order to
show a personalized movie recommendation to the users. The most unique thing about this research work is that the
authors have made movie recommendations using a proper sequence of images which actually describe the movie
story plot. This actually helps for better visuals. The author have also described the graph based recommendation
system, content-based approaches, hybrid recommender systems, collaborative filtering systems, genere correlations
based recommender system, etc. The proposed algorithm has 4 major phases. Initially, social networking website
like Facebook is used to know the user interest. Later, the movie reviews needs to be analysed and the
recommendations needs to be made. Finally, story plot needs to be generated for better visuals.
Md. Akter Hossain, et. al. [12] proposed NERS which is an acronym for neural engine-based recommender system.
The authors have done a successful interaction between 2 datasets carefully. Moreover, the authors stated that the
results of their system are better than the existing systems because they have incorporated the usage of general
dataset as well as the behaviour-based dataset in their system. The authors have used 3 different estimators in order
to evaluate their system against the existing systems.

3. PROPOSED METHODOLGY

We need to perform preprocessing on the dataset and combine the relevant features into a single feature. Later, we
need to convert the text from that particular feature into vectors. Later, we need to find the similarity between the
vectors. Finally, get the recommendations as per the system architecture mentioned below.

1.1. ARCHITECTURE

14954 www.ijariie.com 635

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 1. System Architecture

1.2. DATASET, EXPLORATORY DATA ANALYSIS & PREPROCESSING

The ‘TMDB 5000 Movie Dataset’ is taken into consideration for movie recommendation purpose in this research
work. This dataset is available on kaggle.com. The dataset is composed of 2 CSV files - ‘tmdb_5000_movies.csv’
and ‘tmdb_5000_credits.csv’
The ‘tmdb_5000_movies.csv’ dataset consists of the following attributes:
 ‘budget’: It indicates the budget of the movie.
 ‘genres’: It indicates the genres of the movie like Action, Documentary, etc.
 A movie can have multiple genres.
 ‘homepage’: It indicates the homepage of the movie. It is basically a website link.
 ‘id’: It indicates movie ID.
 ‘keywords’: It indicates the keywords of the movie. Apart from the title of the movie, keywords give a
quick information about the movie.
 ‘original_language’: It indicates whether the movie is originally created in English or other language.
 ‘original_title’: It is nothing but the movie title.
 ‘overview’: It is a short description of the movie.
 ‘popularity’: It is a metric which indicates popularity.
 ‘production_companies’: It consists of the names of companies which has produced the movie.

14954 www.ijariie.com 636

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

 ‘production_countries’: It consists of the names of the countries in which the movie production took place.
 ‘release_date’: It consists of the release date of the movie. The format used is yyyy-mm-dd where ‘yyyy’
indicates year of release, ‘mm’ indicates the month of release, and ‘dd’ indicates the day of release.
 ‘revenue’: It indicates the revenue earned by the movie.
 ‘runtime’: It indicates the runtime of a movie. Runtime basically means the length of the movie.
 ‘spoken_languages’: It consists of the languades spoken in the movie.
 ‘status’: It indicates the status of the movie. For example, a movie can be released or not released which
basically indicates the status of that movie.
 ‘tagline’: It consists of the tagline of the movie.
 ‘title’: It consists of the title of the movie.
 ‘vote_average’: It indicates the average of the votes.
 ‘vote_count’: It indicates the vote count.

Fig. 1. Statistical data about ‘tmdb_5000_movies.csv’ dataset using pandas Dataframe.describe() method

Fig. 2. Glimpse of the ‘tmdb_5000_movies.csv’ dataset using ‘Titanic’ movie

The ‘tmdb_5000_credits.csv’ dataset consists of the following attributes:

14954 www.ijariie.com 637

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

 ‘movie_id’: It indicates the movie ID.

 ‘title’: It indicates the title of the movie.
 ‘cast’: It consists of the cast of the movie. Cast implies the actors and actresses who appear in the movie.
 ‘crew’: It consists of those people who are concerned with the production of the movie.

Fig. 3. Statistical data about ‘tmdb_5000_credits.csv’ dataset using pandas Dataframe.describe() method

Fig. 4. Glimpse of the ‘tmdb_5000_credits.csv’ dataset using ‘Titanic’ movie

The Exploratory Data Analysis (EDA) has been inspired by Heeral Dedhia’s blog on medium.com.

14954 www.ijariie.com 638

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 5. Top Genres

Movies having the genre as Drama are maximum in number as compared to Family movies and Horror movies. A
movie might have multipe genres.

Fig. 6. Actor with highest appearance

The above figure indicates the actors with the highest appearance in the decreasing order.

14954 www.ijariie.com 639

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 7. Directors with highest movies

The above figure indicates the directors with the highest appearance in the decreasing order.

Fig. 8. Runtime versus Number of movies

As the runtime increases, number of movies are increasing. After certain point, as the runtime increases, the number
of movies decreases. There are some exceptions.

14954 www.ijariie.com 640

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 9. Runtime versus Budget

There are a lot of movies with lower budget and falling in the range of runtime 70 to runtime 150.

Fig. 9. Revenue versus Budget

It can be seen from the above figure that low budget movies have low revenue in general.

14954 www.ijariie.com 641

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 10. Correlation Matrix for ‘tmdb_5000_movies.csv’ dataset

From the above correlation matrix, it can be seen that the diagonal is yellow coloured because similarity of
something with itself is always 1.0, i.e., maximum. Moreover, it can be seen that revenue and vote count have more
similarity as compared to budget and vote count.

Preprocessing steps include removing stopwords, combining the first name and the last name into a single name,
removing punctuation marks, lowercasing the text, etc.

Fig. 11. Director, Keywords, Cast and Genres of a movie are combined into a single feature titled ‘combine_feature’
The ‘combine_feature’ attribute needs to be further processed by using some algorithms.

14954 www.ijariie.com 642

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

1.3. ALGORITHMS

We can use CountVectorizer or TfidfVectorizer or Glove or Word2Vec in order to create vectors from the text.
After converting the text into vectors, we need to find the similarity between the vectors. Cosine Similarity or
sigmoid_kernel or some other technique can be used to find the similarity between the vectors.

1. Algorithm 1: Content-based Recommendation using CountVectorizer and Cosine Similarity

In this case, we will use CountVectorizer in order to create vectors from the preprocessed text mentioned in the
‘combine_feature’ attribute.
After getting the vectors, we will find the similarity between the vectors using Cosine Similarity.

2. Algorithm 2: Content-based Recommendation using TfidfVectorizer and Cosine Similarity

In this case, we will use TfidfVectorizer in order to create vectors from the preprocessed text mentioned in the
‘combine_feature’ attribute.
After getting the vectors, we will find the similarity between the vectors using Cosine Similarity.

After getting the recommendations using Algorithm 1 and Algorithm 2, get the common movies from both the
recommendations initially. Later, append the remaining movies to the common movies in an alternate fashion.

4. RESULT AND ANALYSIS

Fig. 12. Recommendations similar to ‘The Dark Knight’ movie using Algorithm 1 and Algorithm 2

14954 www.ijariie.com 643

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 13. Final Recommendations similar to ‘The Dark Knight’ movie

Fig. 14. Recommendations similar to ‘Avatar’ movie using Algorithm 1 and Algorithm 2

14954 www.ijariie.com 644

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 15. Final Recommendations similar to ‘Avatar’ movie

Fig. 16. Recommendations similar to ‘The Godfather’ movie using Algorithm 1 and Algorithm 2

14954 www.ijariie.com 645

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 17. Final Recommendations similar to ‘The Godfather’ movie

Fig. 18. Recommendations similar to ‘Titanic’ movie using Algorithm 1 and Algorithm 2

14954 www.ijariie.com 646

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

Fig. 19. Final Recommendations similar to ‘Titanic’ movie

4. CONCLUSION
We can see from the results that the final recommendations are slightly better than the individual recommendations
of Algorithm 1 and Algorithm 2 mentioned in this research work. Hence, it is always better to manipulate the results
of different algorithms to get the final result which has the advantages of the individual algorithms.

5. REFERENCES

[1] Choi, Sang-Min, Sang-Ki Ko, and Yo-Sub Han. "A movie recommendation algorithm based on genre
correlations." Expert Systems with Applications 39.9 (2012): 8079-8085.
[2] Lekakos, George, and Petros Caravelas. "A hybrid approach for movie recommendation." Multimedia tools and
applications 36.1 (2008): 55-70.
[3] Das, Debashis, Laxman Sahoo, and Sujoy Datta. "A survey on recommendation system." International Journal
of Computer Applications 160.7 (2017).
[4] Zhang, Jiang, et al. "Personalized real-time movie recommendation system: Practical prototype and evaluation."
Tsinghua Science and Technology 25.2 (2019): 180-191.
[5] Rajarajeswari, S., et al. "Movie Recommendation System." Emerging Research in Computing, Information,
Communication and Applications. Springer, Singapore, 2019. 329-340.
[6] Ahmed, Muyeed, Mir Tahsin Imtiaz, and Raiyan Khan. "Movie recommendation system using clustering and
pattern recognition network." 2018 IEEE 8th Annual Computing and Communication Workshop and
Conference (CCWC). IEEE, 2018.
[7] Arora, Gaurav, et al. "Movie recommendation system based on users’ similarity." International Journal of
Computer Science and Mobile Computing 3.4 (2014): 765-770.
[8] Subramaniyaswamy, V., et al. "A personalised movie recommendation system based on collaborative filtering."
International Journal of High Performance Computing and Networking 10.1-2 (2017): 54-63.
[9] Harper, F. Maxwell, and Joseph A. Konstan. "The movielens datasets: History and context." Acm transactions
on interactive intelligent systems (tiis) 5.4 (2015): 1-19.
[10] R. Lavanya, U. Singh and V. Tyagi, "A Comprehensive Survey on Movie Recommendation Systems," 2021
International Conference on Artificial Intelligence and Smart Systems (ICAIS), 2021, pp. 532-536, doi:
10.1109/ICAIS50930.2021.9395759.
[11] N. Immaneni, I. Padmanaban, B. Ramasubramanian and R. Sridhar, "A meta-level hybridization approach to
personalized movie recommendation," 2017 International Conference on Advances in Computing,
Communications and Informatics (ICACCI), 2017, pp. 2193-2200, doi: 10.1109/ICACCI.2017.8126171.

14954 www.ijariie.com 647

Vol-7 Issue-4 2021 IJARIIE-ISSN(O)-2395-4396

[12] M. A. Hossain and M. N. Uddin, "A Neural Engine for Movie Recommendation System," 2018 4th
International Conference on Electrical Engineering and Information & Communication Technology (iCEEiCT),
2018, pp. 443-448, doi: 10.1109/CEEICT.2018.8628128.
[13] Monika D.Rokade ,Dr.Yogesh kumar Sharma,”Deep and machine learning approaches for anomaly-based
intrusion detection of imbalanced network traffic.” IOSR Journal of Engineering (IOSR JEN), ISSN (e): 2250-
3021, ISSN (p): 2278-8719
[14] Monika D.Rokade ,Dr.Yogesh kumar Sharma”MLIDS: A Machine Learning Approach for Intrusion Detection
for Real Time Network Dataset”, 2021 International Conference on Emerging Smart Computing and
Informatics (ESCI), IEEE
[15] Monika D.Rokade, Dr. Yogesh Kumar Sharma. (2020). Identification of Malicious Activity for Network Packet
using Deep Learning. International Journal of Advanced Science and Technology, 29(9s), 2324 - 2331.
[16] 4. Sunil S.Khatal ,Dr.Yogesh kumar Sharma, “Health Care Patient Monitoring using IoT and Machine
Learning.”, IOSR Journal of Engineering (IOSR JEN), ISSN (e): 2250-3021, ISSN (p): 2278-8719
[17] 5. Sunil S.Khatal ,Dr.Yogesh kumar Sharma, “Data Hiding In Audio-Video Using Anti Forensics Technique
ForAuthentication”; IJSRDV4I50349, Volume : 4, Issue : 5
[18] Sunil S.Khatal Dr. Yogesh Kumar Sharma. (2020). Analyzing the role of Heart Disease Prediction System
using IoT and Machine Learning. International Journal of Advanced Science and Technology, 29(9s), 2340 -
2346.

14954 www.ijariie.com 648