“ MOVIE RECOMMENDATION SYSTEM”

A Project Report Submitted

In Partial Fulfillment of the Requirements or the Degree of

Bachelor of Technology (B. Tech)

In

Computer Science & Engineering

Submitted by
Muskan verma (2005050100035)
Piyush Mishra (2005050100038)
Niket Chaurasia (2005050100038)

Under the Supervision of

Mr. Gaurav Tiwari
(Assistant Professor of Computer Science Department)

Allenhouse Institute of Technology

Dr. A.P.J. Abdul Kalam Technical University, UTTAR PRADESH, LUCKNOW
JUNE, 2024
 CERTIFICATE

This is to certified that the Project report entitled “MOVIE RECOMMENDATION SYSTEM”
submitted by Muskan Verma (2005050100035), Piyush Mishra (20050501000038), Niket Chaurasia
(2005050100037) , are Bonafide students of Allenhouse Institute of Technology, Kanpur, Affiliated to Dr.
A.P.J. Abdul Kalam Technical University, Lucknow in partial fulfilment for the award of the Bachelor
of Technology in Computer Science Engineering during the academic year 2023-2024. It is certified that
all corrections/suggestions indicated for internal assessment have been incorporated in the phase 1 project
report deposited in the departmental library. The project work has been approved, as it satisfies the
academic requirements in respect of project work prescribed for the said degree . Project Guide Mr.
Gaurav Tiwari, Professor, Computer Science Engineering Department, Head of the Department Dr.
Sudhir Singh Professor, Computer Science Engineering Department.

Signature
Mr. Gaurav Tiwari
(Assistant Professor
CSE Department, AIT)
Date:

i
 ACKNOWLEDGEMENT

With the guidance and assistance of numerous well-wishers, an endeavor over a

lengthy period of time can be effective. We would like to take this time to let
everyone know how much we appreciate them.

In the beginning, I'd like to express our gratitude to our supervisor, Mr. Gauarv
Tiwari , Assistant Professor, Department of Computer Science & Engineering
Allenhouse Institute Of Technology , for his invaluable support and direction
throughout the project's implementation.

We wish to express our sincere thanks and gratitude to our project guide, Mr.
Gaurav Tiwari , Associate Professor, Department of Computer Science &
Engineering Allenhouse Institute Of Technology , for the stimulating discussions,
in analyzing problems associated with our project work, and for guiding us
throughout the project. Project meetings were highly informative. We express our
warm and sincere thanks for the encouragement, untiring guidance, and
confidence she has shown in us. We are immensely indebted to her for her
valuable guidance throughout our project.

ii
 TABLE OF CONTENT

1. INTRODUCTION 1

1.1 Introduction 1
1.1.1 Natural Language Processing --------------------------------------- 2
1.1.2 Movie Recommendation System ----------------------------------- 3
1.2 Problem Statement 4
1.3 Objectives 4
1.4 Methodology 5
1.4.1 Dataset 6
1.4.2 Flowchart 13
1.4.3 Algorithm 14

2. LITERATURE SURVEY 16

3. SYSTEM DEVELOPMENT 20

3.1 System Configuration 19

3.2 Software Requirement 20
3.3 System Analysis and design 23
3.4 Activity Diagram 19
3.5 Data Flow Diagram 25

4. EXPERIMENT AND RESULT ANALYSIS ------------------------------- 27

4.1 Experiment 26
4.2 Implementation 30
4.3 Method Analysis 35
4.4 Output at various Stages 42

5. CONCLUSIONS 46

5.1Conclusions 46
5.2 Future Scope 47
5.3Application 48
6 .REFERENCES--------------------------------------------------------------------------- 50
iii
 LIST OF ABBREVIATIONS

SHORT FORM MEANINGS

TFIDF = Term Frequency - Inverse Document Frequency

SVM = Support Vector Machine

DT = Decision Tree

GBC = Gradient Boosting Classifier

LR = Logistic Regression

RFC = Random Forest Classifier

CV = Count Vectorizer

FIG = Figure

iv
 LIST OF FIGURES

Figure Page no.

Fig. 1.1: Deep Learning vs Machine Learning vs Artificial 2

Fig. 1.2: Data Pre-Processing 7

Fig. 1.3: Collaborative vs Content based filtering 9

Fig. 1.4: Average Ratings 10

Fig.1. 5: Recommendation avatar 12

Fig. 1.6: Flowchart 13

Fig. 3.1: System architecture for proposed model 21

Fig. 3.2: Tags for movie Avatar 23

Fig. 3.3: Activity Diagram 24

Fig. 3.5: Avatar movie example 25

Fig. 3.6: Data flow diagram 30

Fig. 4.1: Result 37

Fig. 4.2: Content Based recommendation 38

Fig. 4.3: Collaborative Filtering 40

Fig. 4.4: Hybrid Filtering 42

Fig. 4.5: Code implementation 45

Fig. 4.6: Code implementation 46

v
 LIST OF GRAPHS

Page no.

Graph 1: Category Distribution 6

Graph 2: Genre Distribution 8
Graph 3: Comparison of F-Measure 11
Graph 4: Frequency of subject of the news 28
Graph 5: Feature weight of Avatar 29

vi
 LIST OF TABLES

Table Page No.

Table 1: Literature Review 16

Table 2: Categories 27

Table 3: Information of Movie Avtaar 29

Table 4: Pros-Cons of content-based filtering 36

Table 5: User Based CF 38

Table 6: Item based CF 39

Table 7: Pro-Cons collaborative filtering 40

vii
 ABSTRACT

In this hustling world, entertainment is a necessity for each one of us to

refresh our mood and energy. Entertainment regains our confidence for
work and we can work more enthusiastically. For revitalizing ourselves, we
can listen to our preferred music or can watch movies of our choice. For
watching favourable movies online we can utilize movie recommendation
systems, which are more reliable, since searching of preferred movies will
require more and more time which one cannot afford to waste. In this
paper, to improve the quality of a movie recommendation system, a Hybrid
approach by combining content based filtering and collaborative filtering,
using Support Vector Machine as a classifier and genetic algorithm is
presented in the proposed methodology and comparative results have been
shown which depicts that the proposed approach shows an improvement in
the accuracy, quality and scalability of the movie recommendation system
than the pure approaches in three different datasets. Hybrid approach helps
to get the advantages from both the approaches as well as tries to eliminate
the drawbacks of both methods.

viii
 CHAPTER-1
INTRODUCTION

1.1) Introduction

The objective of this project is to build a movie recommendation system using

MERN stack ( MangoDB , Express.js , React , Node.js ) integrated with
machine learning algorithms . the system aims to provide personalized movie
recommendation to users based on their preferences and viewing history.

The study of statistical models and methods used by computers to do certain

tasks devoid of explicit instructions and in favour of patterns and inference is
known as machine learning (ML). It's thought to be a part of artificial
intelligence. Without being explicitly told to do so, machine learning
algorithms create a mathematical model from sample data, or "training data,"
in order to produce conclusions or predictions. There are many similarities
between machine learning and computational statistics, which focuses on
computer-aided prediction. Machine learning benefits from the
methodologies, theories, and fields of application created through the study
of mathematical optimization.

A recommendation system, sometimes known as a recommendation engine, is

a paradigm for information filtering that aims to anticipate user preferences

and offer suggestions in accordance with these preferences. These

technologies are now widely used in a variety of industries, including those

that deal with utilities, books, music, movies, television, apparel, and

restaurants. These systems gather data on a user's preferences and conduct,

which they then employ to enhance their suggestions going forward.

Movies are a fundamental aspect of life. There are many various kinds of
movies, such as those meant for amusement, those meant for teaching,
1
children's animation movies, horror movies, and action movies. Movies'
genres, such as comedy, thriller, animation, action, etc., make it simple to
distinguish between them. Another approach to differentiate between movies
is to look at their release year, language, director, etc. When watching films
online, there are many to choose from in our list of top picks. We can use
movie recommendation systems to find films based on favoured films among
all of these other movie genres, saving us the hassle of having to spend a lot
of time looking for favourite films. As a result, it is essential that the system
for suggestion films to us is very trustworthy and gives us recommendation
for the films that are either most similar to or identical to our tastes.

Figure:1.1

1.1.1) Natural Language Processing (NLP)

A branch of computer science and artificial intelligence known as "natural

language processing," or NLP, studies how computers interact with human
(natural) languages with the goal of effectively teaching computers to analyse
massive volumes of natural language data. The study of how computers
interact with human (natural) languages is known as natural language
processing (NLP), and it is a branch of linguistics, computer science,
2
information engineering, and artificial intelligence. Its primary goal is to train
computer programmers to evaluate and interpret vast amounts of natural
language.

1.1.2) Movie Recommendation System

Recommendation systems are being used by a lot of businesses to improve

customer interaction and the purchasing experience. The most significant
advantages of recommendation systems are client happiness and income. A
very effective and crucial mechanism is the movie recommendation system.
However, because of the limitations with a pure collaborative method,
scalability concerns and poor recommendation quality also affect movie
recommendation systems.

This feature certainly intrigues me. As a result, the main duty of a

recommender system is to provide the user with the most useful
recommendations. While Amazon, Flipkart, and Netflix utilize
recommendation algorithms for product recommendations, Amazon Prime
and YouTube use them for movie recommendations.

Any action you undertake on these websites is being tracked by a system,

which then makes suggestions for goods or products that you are very likely
to find interesting. This study looks at movie recommendations, the logic
behind them, as well as more traditional movie recommendation systems and
a solution for an AI-based customized movie recommendation system.

From a business perspective, user engagement is higher the more relevant

products they discover on the site. Increasing platform income is a common
outcome of this. From a business perspective, user engagement is higher the
more relevant products they discover on the site. Increasing platform income
is a common outcome of this. Various sources claim that as much as 35–40%
of the revenue of internet behemoths comes from only referrals.

3
1.2) Problem Statement

The goal is to develop a movie recommendation system that can provide users
with tailored movie suggestions based on their tastes in films. Based on a user's
historical movie ratings and preferences, as well as suggestions of comparable
films seen by other users with similar likes, the system should be able to forecast
with accuracy which films the user will likely appreciate. The system should also
be able to scale easily and handle enormous volumes of data with efficiency. By
suggesting films that the user is likely to enjoy, the system hopes to improve user
experience and increase user engagement and retention.

1.3) Objective

Creating a personalized system that can make movie suggestions to users

based on their prior movie choices is the main goal of a movie
recommendation system project. Utilizing machine learning algorithms, the
system will analyse user data and produce recommendations that are relevant
to their interests. The following objectives are the focus of the project:

• Providing accurate and customized movie recommendations to users

based on their past behaviour and preferences.
• Analysing user data utilizing machine learning algorithms to produce
movie recommendations.
• Considering multiple factors such as user movie ratings, movie genre,
movie director, and similar movies watched by other users with
similar preferences to produce recommendations.
• Designing a scalable and effective system capable of processing large
amounts of data and delivering rapid recommendations.
• Enhancing user engagement and retention by providing a smooth and
enjoyable movie recommendation experience.

4
1.4) Methodology

1.4.1) Dataset

A data set (or dataset) is a collection of data. In the case of tabular data,
a data set corresponds to one or more database tables, where every
column of a table represents a particular variable, and each row
corresponds to a given record of the data set in question . To develop a
movie recommendation system, you can follow these general steps:

1. Collect data on movies, such as their title, genre, director, actors,

release date, and ratings. You can use public datasets such as the IMDb
dataset, or collect your own data through web scraping or API calls.

Movie dataset has

• Movie Id – once the recommendation is done, we get a list of all similar

movieId and get the title for each movie from this dataset.

• Genres – which is not required for this filtering approach.

• Budget – money spent in making money

• Original language – initial language in which movie is made

• Production Companies- it tells company that made the movie

• Cast – actors and actresses the acted in the movie

• Keywords – words that describe the movie and can be used to indentify

the movie.

5
Category Distribution:

It shows which category or genre movie is being seen maximum number

of people

Graph :1.1

2. Clean and preprocess the data by removing duplicates, missing values,

and irrelevant columns. You can also use feature engineering
techniques to create new features that capture the characteristics of the
movies. To remove duplicates in machine learning, the following
steps can be taken:

• Locate and identify duplicate records in the dataset using a unique

identifier or a combination of attributes.

• Choose a criterion for selecting one record from each group of

duplicates, such as selecting the first occurrence or the one with the

highest or lowest value of a particular attribute.

• Eliminate the duplicate records from the dataset, retaining only the
chosen record for each group of duplicates.
6
 Figure:1.2

3. Perform exploratory data analysis to gain insights into the data, such as
the distribution of ratings, the most popular genres, and the correlations
between different features. Examining and interpreting data to derive
important insights and conclusions is the process of data analysis. It
entails analysing huge datasets using a variety of statistical and
computational tools to find patterns, trends, and relationships.

Using descriptive statistics, we may enumerate and explain the key

characteristics of a dataset. These include measures of variability like
standard deviation and range as well as measures of central tendency like
mean, median, and mode.
Using inferential statistics, it is possible to predict and infer information
about a broader population from a sample of data. Confidence intervals
and hypothesis testing are some of these methods.

7
 Graph : 1.2
4. Choose a machine learning algorithm to build the recommendation
system. Some popular algorithms for recommendation systems include
collaborative filtering, content-based filtering, and hybrid filtering. Movie
recommendation systems mainly use three types of algorithms to provide
personalized recommendations to users:

• Content-Based Filtering: This algorithm analyses a user's previous movie

ratings or preferences and recommends new movies based on the content
of the movies. It identifies patterns and similarities between the movies
and suggests new movies that have similar characteristics to the ones that
the user has already watched and liked.

8
• Collaborative Filtering: This algorithm recommends movies based on the
user's behaviour and patterns in the past. It analyses the user's movie
ratings and preferences, as well as those of other users with similar tastes.
Based on this analysis, it identifies movies that the user may be interested
in and recommends them.

Figure : 1.3

• Hybrid Recommendation: This algorithm combines both content-based

and collaborative filtering algorithms to provide a more accurate and
personalized recommendation. It leverages the strengths of both
algorithms to overcome their individual weaknesses and produce more
relevant and effective recommendations.

• Overall, these three types of algorithms are widely used in the development
of movie recommendation systems to enhance user engagement and
satisfaction by providing tailored and relevant recommendations.

5. Train the model on the movie data, using techniques such as matrix
factorization, deep learning, or clustering. Training a movie
recommendation model involves feeding it with a dataset of movie ratings
and other relevant information, such as movie genres, actors, directors,
and release years.

9
 Figure : 1.4

6. Evaluate the performance of the model using metrics such as

precision, recall, and F1 score. You can also use techniques such as
A/B testing or user studies to evaluate the user satisfaction with the
recommendations. The evaluation of a movie recommendation
system's performance is essential to ensure that it is delivering
accurate and relevant recommendations to its users.

• Mean Absolute Error (MAE): This metric measures the average

absolute difference between the predicted and actual ratings. The
lower the MAE, the better the system's performance.

10
 • Root Mean Square Error (RMSE): This metric measures the square
root of the average squared difference between the predicted and actual
ratings. As with MAE, a lower RMSE indicates better performance.

• Precision and Recall: These are classification metrics used to

evaluate the accuracy of binary recommendation systems that predict
whether a user will like a movie or not. Precision measures the
proportion of recommended movies that the user actually liked, while
recall measures the proportion of movies that the user liked that were
recommended by the system.

Figure :1.3

7. Deploy the recommendation system as a web application or API, where

users can input their preferences and receive personalized movie
recommendations. Model deployment is the action of implementing
machine learning models. This makes the model's predictions accessible
to users, developers, or systems, allowing them to interact with their
application (such as identify a face in an image) or make business
decisions based on data.

11
 Figure : 1.5

When developing a methodology for a movie recommendation system,

it's important to consider factors such as user demographics, user
feedback, and the diversity of the recommended movies. Additionally,
it's crucial to ensure that the system is transparent, interpretable, and
respects user privacy.

12
1.4.2) Flowchart:

Figure : 1.6

1.4.3) Algorithm for Th Proposed System

Step 1: Pre-processing

Removing the repeated attributes data.

Clean the text by removing punctuation, stopwords, and lowercasing the
text.

Split the dataset into training and testing sets.

Step 2 : Count Vectorization

▪ Use Count Vectorizer from sklearn library to convert the text data
into numerical data
▪ Create a document-term matrix that represents the frequency of
each word in each document

13
 ▪ Use the training set to fit the Count Vectorizer and transform the data
▪ Use the testing set to transform the data.

Step 3 : TFIDF Vectorization

▪ Use TF-IDF Vectorizer from sklearn library to convert the text data
into numerical data
▪ Create a document-term matrix that represents the importance of
each word in each document
▪ Use the training set to fit the TF-IDF Vectorizer and transform the data
▪ Use the testing set to transform the data

Step 4 : Training the Models

▪
Use the transformed data from Count Vectorizer and Tfidf
Vectorizer to train different models such as Naive Bayes, Logistic
Regression, Support Vector Machines (SVM), Random Forest, etc.

▪
Use the training set to fit the models
▪
Use the testing set to predict the labels of the news articles

▪
Calculate the accuracy score of each model using the predicted
labels and the actual labels

Step 5: Confusion Matrix

▪
Create a confusion matrix for each model to evaluate its performance
▪
The confusion matrix shows the number of true positives, true
negatives, false positives, and false negatives
▪
Use the confusion matrix to calculate metrics such as precision,
recall, and F1-score

Step 6 : Accuracy

▪
Calculate the accuracy of each model using the predicted labels
and the actual labels

14
▪
Accuracy is how close a given set of measure are to their true value

15
 CHAPTER-2
LITERATURE SURVEY

Over the years, many recommendation systems have been developed using
either collaborative, content based or hybrid filtering methods. These systems
have been implemented using various big data and machine learning
algorithms.

[1] Unnathi Bhandari, D. Garg, M. A. Maarof, and R. A. Rashid's paper ": A

survey" is a survey; it contains no experiments or findings. Instead, the study
presents a thorough analysis of the pros and cons of the many movie
recommendation techniques suggested in the literature, as well as the datasets
that were utilised to test them. The methods utilised by various research in
terms of feature selection, feature extraction, classification algorithms, and
assessment metrics are analysed and compared by the authors. They also
emphasise the difficulties and potential avenues for future study in the area of
movie recommendation systems.

[2] In the paper, MovieREC—a recommender system for movies—is

introduced. It enables a user to choose from a predetermined set of criteria
and then suggests a list of films for him based on the cumulative weight of the
various attributes and the K-means algorithm. Author selects K initial
centroids in the K-means clustering algorithm, where K is the required
number of clusters. Each point is subsequently assigned to the cluster's
centroid, which has the closest mean. Then, using the points allocated to each
cluster, we update the centroid of each cluster.

16
[3] The author use content-based filtering, which is determined by the item's
description and the user's preference profile. In CBF, we employ keywords in
place of the user's profile to represent an item's preferred likes and dislikes. In
other words, CBF algorithms promote products that were previously liked or
products that are related to those products. It looks at previously rated things and
suggests the best item that matches.

[4] The author contrasts different methods for creating a movie recommendation
system. Hybrid recommender systems frequently combine these methods. An
earlier study by Eyjolfsdottir et al. for the suggestion of films through
MOVIEGEN had some shortcomings, including the time- consuming set of
questions it asks consumers. However, it was not user- friendly due to the fact
that it turned out to be somewhat stressful.

[5] With these drawbacks in mind, authors created MovieREC, a movie

recommendation system that makes movie suggestions to consumers based on
the data they submit. In the current study, a user has the ability to choose from a
variety of variables, such as actor, director, genre, year, and rating, among others.
Based on the preferences of users' prior visited histories, we forecast the users'
selections. The system was created in PHP and at the moment only offers a
straightforward console-based user interface.

[6] Author selects K initial centroids in the K-means clustering algorithm, where
K is the required number of clusters. Each point is subsequently assigned to the
cluster's centroid, which has the closest mean. After that, based on the points
assigned to each cluster, the author updates the centroid of each cluster. Once the
cluster centre (centroid) had not changed, the procedure was repeated. Last but
not least, the objective of this algorithm is to minimise an objective function, in
this case a squared error function.

17
[Ref. Author(s) Published By (IEEE, Pros and cons
No.] Elsevier, Sprin ger)

Deepati Garg, Movie Recommendation The model doesn't need any data
[1] Unnati Bhandari, System Using about other users, since the
Ching Sen Collaborative Filtering. recommendations are specific to
this user. This makes it easier to
scale to a large number of users.

Accuracy is low as compared to

other.

Nitasha Soni, Machine Learning Based The model can capture the
[2] Krishan Kumar, Movie Recommendation specific interests of a user, and
Ashish Sharma, System can recommend niche items that
Aman Yadav very few other users are interested
in.

It was trained only one model

with accuracy of 80.035

[3] Narendra Kumar Movie Recommendation Content-based filtering uses

Rao, System using Machine similarities in products, services,
Nagendra Learning or content features, as well as
Panini Challa information accumulated about
the user to make
recommendations.

P. Karthikeyan, Review of Movie Since the feature representation of

C. Tejaswani Recommendation System the items are hand-engineered to
some extent, this technique
requires a lot of domain
knowledge. Therefore, the model
[4] can only be as good as the hand-
engineered features.

[5] Shourya Chawla, Machine The model can only make

Sumita Gupta. Recommendation Models recommendations based on
Rana Majumdar using Machine Learning existing interests of the user. In
other words, the model has
limited ability to expand on the
users' existing interests.

18
 Kevin Andrews, Web based movie Collaborative filtering relies on
[6] Lakshmi recommendation system the preferences of similar users to
Narayan , K using content based offer recommendations to a
Balasubramanian filtering particular user.
,M S Josephine

Table : 2.1

19
 CHAPTER-3
SYSTEM DEVELOPMENT

3.1) System Configuration

Common hardware can be used to run this project. We used an Intel I5 CPU with 8
GB of RAM, a 2 GB Nvidia graphics processor, and 2 cores with respective clock
speeds of 1.7 GHz and 2.1 GHz to complete the project. Predictions may be made
and accuracy can be assessed in a couple of seconds during the test phase, which
follows the training phase and lasts for approximately 10-15 minutes.

3.2) Software Requirements

Distribution of anacondas:

Python is a free and open-source programming language that may be used for
scientific computing (data science, machine learning), and Anaconda is a
distribution of it that aims to simplify the package management system and
deployment (for things like apps, big data processing, predictive analytics, etc.).
Package versioning is managed by a system called Conda. The Anaconda
distribution comes includes data science packages that work with Windows, Linux,
and MacOS.3

3.2.1) Python Libraries

Scikit-Learn (sklearn): The classification, regression, and clustering techniques

in this collection include support vector machines, random forests, gradient
boosting, k-means, and DBSCAN, among others. It is made to function perfectly
with Python’s NumPy and SciPy scientific and numerical libraries.

NumPy: A well-liked all-purpose array processing package is NumPy. In addition to

tools for working with these arrays, it offers a high-performance multidimensional

20
array object. It is a foundational Python package for scientific computing.

Pandas: Another popular Python package in data science is Pandas. It offers user-
friendly, high-performance structures and tools for data analysis. A Data Frame is
a 2D table object that may be stored in memory in Pandas, as opposed to NumPy,
which offers objects for multidimensional arrays.

Flask: The WSGI (Web Server Gateway Interface) web application framework
Flask is compact. With the potential to scale up to complicated applications, it is
made to set up quickly and effortlessly. It started out as a straightforward wrapper
for Werkzeug and Jinja but has since grown to be one of the most well-liked
Python web application frameworks.

Matplotlib: A plotting library for the Python programming language is called

Matplotlib. It offers a selection of static, animated, and interactive Python
visualizations.

In a movie recommendation system project, these libraries are essential for

carrying out a number of tasks, including data processing, data visualization,
machine learning, and web application development. Developers can create
effective and precise movie recommendation systems that offer customers
personalized movie recommendations by utilizing these libraries.

21
3.3) System Analysis & Design

figure:3.1

System architecture for proposed model

A common strategy in recommender systems is content-based recommendation,

which aims to suggest products that are comparable to those that a consumer has
previously shown an interest in. A content-based recommender system's primary
objective is to determine how similar products are to one another. There are
several ways to model items, with the Vector Space Model being one of the most
widely used.

22
The TF-IDF is used to extract keywords from items and determine their weights in
the Vector Space Model. Let ki be the ith keyword and w ith be the weight of ki
for the provided item , dj. So , a series of weights can be used to indicates the
content of dj: content(dj) is equal to “ w1j,w2j,....”

Equation:1

Based on their history of liked items, a user's preference vector, Content- Based
Profile(u), can be constructed to model their preferences. The following definition
of Content-Based Profile(u) can be used if N(u) is the collection of items that user
u has liked:

Equation :2

This makes it possible for the content-based recommender system to provide the
user with recommendations based on their preferences and similarities among
things. N(u) is the previous user that u loved. Given each user u and an item d, the
similarity between the content vector Content(.) and the content preference vector
Content Based Profile of all users indicates how the user feels about the item:

Equation :3

23
 Figure:3.2
Tags for movie Avatar

3.4) Activity Diagram

Activity diagrams are visual depictions of workflows with choice, iteration, and
concurrency supported by activities and actions. Similar to the other four
diagrams, activity diagrams serve similar fundamental goals. It captures the
system's dynamic behaviour . The message flow from one item to another is
depicted using the other four diagrams, whereas the message flow from one
activity to another is depicted using the activity diagram.

24
 Figure : 3.3

The user is given a list of recommended films after logging in using the user id,
which is available in the csv file and ranges from 1-5000. After that, each movie in
the test set is classified, which in our case involves assigning a genre to each
movie. Since we now know the appropriate movie genre, the following part will
look at the appropriate and erroneous categorizations and utilize metrics to judge
the advancement.

25
 Figure :3.4

3.5) Data Flow Diagram

Figure:3.5

Initially, it is best to load the data sets required to build a model . This project
requires the use of the files movies.csv, rats.csv, and users.csv. Each data set can
be found on the Kaggle.com website. This project's material essentially creates two
models.

26
 CHAPTER-4
EXPERIMENT AND RESULT ANALYSIS

4.1) EXPERIMENT

Initially, it is best to load the data sets required to build a model.This project requires
the use of the files movies.csv, rats.csv, and users.csv. Each data set can be found on
the Kaggle.com website. This project's material essentially creates two models.

4.1.1) Data Set

From the perspective of a recommender system, a movie can be described by a

number of characteristics, including genres, actors, directors, and so on.

• Director: IMDb was used to obtain the director's information; while most

films only have one director, some do have two or more.

• Actors: Large casts are common in films, however the vast majority of them are

detrimental to the recommender system and have negative effects. As a result,

there are just three notable actors in the film. They come from IMDb as well.

• Keyword: We use LSI to extract keywords from the Wikipedia plot with help

from our pals at VionLabs.

• Release Year: The information is from IMDb and represents the year

the movie was released.

27
4.1.2) Category

For a movie, we'll divide the films into 23 categories based on the common genres. Here is
a list of the categories we used. Each movie-based document in the case is represented by
one of the eight features specified in Section. The video is represented using a vector space
model, and each feature of the document includes the word "movie."

Table:4.1

The genre is often used as a vector in other content-based recommender systems to

estimate how similar two things are. This is only one element of the movie; there are
many others, such as the actor and the setting. As a result, we add new features, some
of which are really distinctive because they were found via our own research.

The reasons, though, are why we didn't just add features together to calculate TF-
IDF. A natural trait that can be used to classify something is the genre.

4.1.3) Document

As we previously explained, the document in this instance is a video that has a lot

28
of features. In the experiment, a vector space model will represent the movie. We
described the characteristics that the movie is modelled on. The format of the
vector space model is as follows:
Movie Model = [ Writers , Performers , Keywords , Year of Release , violin
themes , Languages , Places , violin Scenes ]

Graph : 4.1

The vector is typically quite long because there can be numerous directors and
actors in a single movie. Here, we use the movie Avatar to show how the model
works. Our TF-IIDF-DC calculated that The Dark Knight has 80 distinct features;
this figure highlights the importance of each feature. From this vantage point, it is
clear that a similar distribution of a film's features denotes a film's similarity to
another

29
 Graph : 4.2

Table : 4.2

4.1.4) Result

Feature to cinema in this context refers to the phrase document. The video format
of the vector space model, which may be used to assess similarity, can be easily
converted. Thanks to preceding calculations, each movie in the database may be
represented by a vector. The cosine similarity approach was then used to calculate
how similar one movie is to the others.

30
 Figure:4.1

4.2) Implementation

The System Make Use Different Algorithms and Methods for the implementation of
Content Approach

4.2.1) Cosine Similarity:

The similarity of two non-zero vectors in an inner product space is measured by the
cosine of the angle between them. Cosine similarity is a statistic that is used to assess
how similar two documents are, regardless of the size of the documents. The cosine of
the angle made by two vectors projected onto a multidimensional space is computed.
Even if the two comparable documents are separated by a significant Euclidean distance
due to the size of the documents, the cosine similarity is advantageous since it enhances
the possibility that the two comparable documents will be oriented closer together. As
the angle gets smaller, the cosine similarity gets stronger.

31
 Equation : 4

4.2.2) Singular Value Decomposition (SVD):

With singular vectors v1, v2,..., vr and corresponding singular values 1, 2,..., r, let
A be a n*d matrix. The left singular vectors are then ui = (1/i)Avi, where i = 1,
2,..., r, and according to Theorem 1.5, A may be broken down into a sum of rank
one matrices.

Equation : 5

First, we provide a straightforward lemma that states two matrices A and B are
equivalent if Av = Bv for all v. The lemma argues that a matrix A can be thought
of as a transformation that translates vector v onto Av in the abstract.

4.2.3) Manhattan Distance :

The Manhattan distance metric measures the distance between two points as the
sum of the absolute differences between their Cartesian coordinates. The sum of
the discrepancies between the x- and y-coordinates can be used to express it. When
p1 and p2 are situated in a plane at (x1, y1) and (x2, y2), respectively,

32
 Equaton:6

4.2.4) Euclidean Distance:

The Euclidean distance between two points in either flat or three-dimensional

space determines the length of a segment connecting them. It is the method that
best demonstrates a distance between two locations. How far apart two points are
can be calculated using the Pythagorean Theorem.
Formula: The euclidean distance in two dimensions between the points (x1,y1)
(x1,y1) and (x2,y2) (x2,y2).

Equation :7

4.2.5) Jaccard Similarity:

The Jaccard index, sometimes referred to as Intersection over Union and the
Jaccard similarity coefficient, is a statistic for evaluating the similarity and variety
of sample sets. The Jaccard coefficient measures the similarity between finite
sample sets by dividing the size of the intersection by the size of the union of the
sample sets. The Jaccard index, often known as the Jaccard similarity coefficient,
is a statistic for assessing the diversity and similarity of sample sets.

33
 Equation:8

Comparison Between Cosine and Manhattan Distance

Cosine similarity or Euclidean distance can be used to calculate the distance

between two vectors in a vector space. Because it's not always evident what we
mean when we talk about the distance between two vectors, as we'll see in a
moment, it's imperative that we be explicit about what we mean.
In a 2D vector space, three distinct points—blue, red, and green—are located. We
could ask ourselves which pair or pairs of points are closer to one another. As we
go, we expect that the answer will include a specific pair (or pairs) of points:

If just one combination is the closest, the answer can either be (red, green), (blue,
red), or (blue, green). If two pairs are the closest, then three sets are possible; these
sets match all two-element combinations of the three pairs.

There is only one set that could possibly contain all three couples if they are all
equally near . This indicates that the closest set One of the seven potential sets is a
pair or pairs of points. Then, how can we decide which of the seven potential
solutions is correct? To do this, we must first choose a technique for gauging
distances. Using a ruler and two points, we may measure the reading to determine
which response is correct . If we do this for all possible pairs, we can generate a
list of measurements for pair-wise distances. The table can then be sorted
ascendingly to reveal the pairwise pairing of points with the least distances.

In this instance, the pair (red, green) makes up the set with the shortest distance.
Thus, we can claim that the shortest Euclidean distance between the red and green

34
points in our collection is the distance measured by a ruler between them. We can
also use an entirely different but equally acceptable method to get the distances
between the identical places. Let's imagine that we are looking at the points from
within the plane, specifically from its inception, as opposed to the top of the plane
or from a bird's eye view. This allows us to depict with an arrow the direction that
we consider when analyzing each point: Regardless of how far apart the points.
From our perspective point, it doesn't really matter how far the points are from the
origin. Actually, without leaving the plane and entering the third dimension, we
are unable to understand that. When viewed from the origin, all of the points
appear to be on the same horizon; their only difference is the path they take in
relation to a reference axis.

Choosing the metric to employ relies on the specific activity that needs to be
completed:

Both metrics are helpful for various tasks, such preliminary data analysis, because
they each make it possible to glean particular insights about the structure of the
data. Euclidean distances usually function better when applied to others, such as
text classification.

The retrieval of the texts that are most similar to a given document is one example
of a more comprehensive application where cosine similarity performs better. The
challenge is in comprehending all methods and learning the heuristics associated
with their use, as is frequently the case with machine learning. One discovers this
by trial and error.

35
4.3) Method Analysis

Recommender systems have grown in popularity as a research topic in recent

years. Many academics have offered a variety of different recommendation tactics.
The most well-known classification of these tactics is

• Content-based recommendations
• Collaborative filtering is advised.
• Recommendation-for hybrids.

4.3.1) Content Based Recommendation System

Based on parameters for movies like genre, director, description, actors, etc., it
provides recommendations for users. A user might love a movie or television show
similar to one they already enjoyed, according to the logic behind this type of
suggestion system.

Many recommender systems begin by modelling the item with keywords. But
extracting keywords from a piece of content can be difficult, especially in the
media sector where it can be difficult to extract text keywords from videos. There
are primarily two methods for resolving this kind of problem. In the first, users can
tag the items, while in the second, experts are involved. Jinni and Pandora,
respectively, are the exemplary expert tagged systems for music and movies. As
an example, consider Jinni, whose researchers identified over 900 tags as "movie
genes" and permitted movie industry pros to generate tags for them. A number of
criteria, including "movie genre," "story," "time," "place," and "cast," apply to
these keywords.

36
ADVANTAGES DISADVANTAGES

• Since the • This technique needs a lot

recommendations are of domain knowledge
particular to this person, because the feature
the model doesn't require representation of the
any information about items is somewhat hand-
other users. This makes engineered. The quality
scaling to a huge user of the model is therefore
base simpler. limited to the hand-
engineered elements.

• The model may identify a • Only recommendations based

user's precise preferences and on the user's current interests
recommend specialised can be made by the model. In
products that only a small other words, the model has
number of other users are little capacity to further
likely to be interested in.. develop the interests of the
users..

Table : 4.3

37
 Figure : 4.2

4.3.2) Collaborative Recommendation System

It matches people with similar interests and gives recommendations based on their
preferences. Sam and Robin, two examples, who favour the movie A, B, C, and D,
respectively. Sam would recommend the films A and B to Robin because C and D
are also favourites of Sam's. Collaborative filtering does not use metadata to
produce suggestions.

I will be focusing on content-based recommendation systems for this project since

I think that using metadata like "genres," "actor," and "overview/plot" will provide
us a lot of insight on understanding users' interests and help recommend films or
TV episodes in line with this. It facilitates the discovery of user preferences.

38
 Figure : 4.3

4.3.2.1) User-based collaborative filtering:

It is presumed that a user will like things that other users with similar tastes will
also like. Thus, the first step in user-based collaborative filtering is identifying
users with similar likes. When users favour related things, this is referred to as
collaborative filtering. In other words, given user u and user v, N(u) and N(v) are
the things set liked by u and v, respectively. As a result, it is simple to determine
how similar u and v are:

Table:4.4

Imagine that we wish to give our friend Stanley a movie recommendation. We

could infer that people who are similar will have similar tastes.

39
4.3.2.2) Item-based collaborative filtering:

It is presumed that a user will like things that other users with similar tastes will
also like. Thus, the first step in user-based collaborative filtering is identifying
users with similar likes. When users favour related things, this is referred to as
collaborative filtering. In other words, given user u and user v, N(u) and N(v) are
the things set liked by u and v, respectively. Thus, it is simple to summarise how
similar u and v are.

An illustration of an item-based CF recommendation is the table. We may infer

that Item A and Item C are comparable since those who like Item A also like Item
C, according to the interest history of all the users for Item A. Since User C enjoys
Item A, it stands to reason that she could also enjoy Item C.

Table : 4.5

Collaborative filtering with user and item based criteria was seen. In the first, the
emphasis is on populating a user-item matrix and making recommendations based
on users who are more like the active user. IB-CF, on the other hand, fills out a
matrix of related objects and makes recommendations.

Although it is challenging to cover all of these topics succinctly, doing so is the

first step in learning more about RecSys

40
Advantages Disadvantages
Due to the fact that the The prediction of the model for a
embeddings are automatically certain (user, item) pair is
taught, we don't require domain represented by the dot product of
expertise. the related embeddings. As a
result, the system cannot embed
or use an item to query the model
if it is not detected during
training.
The model's users might discover By averaging the embeddings of
new hobbies. The machine objects from the same category,
learning algorithm may not be from the same uploader (in
aware of the user's interest in a YouTube), and so forth, the
certain item in a catalogue, but the system can approximate its
model may nevertheless embedding if it lacks interactions.
recommend it because there may
be other users who have the same
interest.
The system may, to a certain Side features are any features that
extent, train a matrix extend past the query or item ID.
factorization model solely on the The user's age or country may be
feedback matrix. The system side variables for movie
doesn't specifically need suggestions. The model's quality
contextual characteristics. increases when accessible side
Actually, it is possible to use any features are added.
of a number of candidate
generators.

Table:4.6

4.3.3) Hybrid Recommendation System

The popularity of hybrid recommender systems is growing right now. According

to recent studies, collaborative filtering and content-based filtering can work better
together. Hybrid recommender systems can be implemented in a variety of ways:
easily add CF capability to a CB technique and aggregate the results of CF and CB
recommendations.

41
There are seven ways to hybridise:

• Weighted : Add the results of the various recommender component

scores.
• Switching : Select ways by alternating between various
recommender components.
• Mixed : Display the outcomes of multiple systems' recommendations.
• Combining features : Taking features from several sources and
combining them into one input.
• Feature Augmentation : Compute features using a single
recommender and go on to the next stage using the results.
• Cascade : Use a recommender technique to generate a rough result, then
recommend it on top of the prior result.
• Meta-level : Input another recommender approach with the model
produced by one recommender.

Each approach has advantages and disadvantages, and the results change based on
the dataset. The approach might not be suitable for all problems because of the
algorithm's inherent constraints. For instance, it is difficult to automate feature
extraction from media data using a content-based filtering strategy. Additionally,
the diversity is not as good because the recommendation only contains products
that the customer has already selected.

Recommending to users who never make decisions is exceedingly difficult.

Collaborative filtering techniques somewhat minimise the previously mentioned
drawback.

However, because CF relies so much on past data, there are issues with cold starts
and sparsity. Due to cold start challenges that involve both new user and new item
issues, collaborative filtering, which is based on the similarity between the things
selected by users, finds it challenging to recommend a new item that has never
been recommended before.

42
4.4) Output at various stages

• We obtained our dataset from Kaggle, which also contains the 5000 films

listed on IMDb and IMDb. I've shown the dataset we're utilising in its

initial form in the image below. I demonstrated it using Python's pandas

package.From the perspective of a recommender system, a movie can be

described by a number of characteristics, including genres, actors,

directors, and so on.

• Director : The director's information is taken from IMDb; most films only

have one director, but others have two or more.

• Actors/cast: Large casts are common in films, however the vast majority

of them are detrimental to the recommender system and have negative

effects. As a result, there are just three notable actors in the film..

• Keyword: We use LSI to extract keywords from the Wikipedia plot with

help from our pals at VionLabs.

• Release Year: This is the film's release year, and the information comes

from IMDb.

• Genres : This shows the type of film comedy , thriller , keywords etc.

43
 Figure :4.4

• Category:

For a movie, we'll divide the films into 23 categories based on the common
genres. Here is a list of the categories we used. Each movie- based document in
the case is represented by one of the eight features specified in Section. The
video is represented using a vector space model, and each feature of the
document includes the word "movie."

Table : 4.7

44
Figure : 4.5

Figure : 4.6

45
Figure : 4.7

Figure :4.8

46
 CHAPTER-5
CONCLUSION

5.1) Conclusion

In this project, content-based filtering and collaborative filtering are combined

using Singular Value Decomposition (SVD) as a classifier and Cosine Similarity
as the recommended approach. This tactic seeks to improve the effectiveness,
superiority, and scalability of movie recommendation systems. On three separate
movie datasets, existing pure algorithms and the hybrid technique are applied, and
the outcomes are compared. Comparative results demonstrate that the suggested
approach outperforms pure alternatives in terms of the accuracy, quality, and
scalability of the movie recommendation system. The proposed solution is quicker
when comparing the computation times of the three pure techniques.

These results suggest that a range of classifiers may be used with comparable
success rates and that machine learning techniques can be highly effective at
spotting bogus news. The models must be evaluated using different metrics, such
as precision, recall, and F1-score, in addition to factors like interpretability,
scalability, and computing requirements. It is important to keep in mind that
accuracy is only one statistic. It may also be useful to look into different feature
extraction and selection methods, classifier types, and ensemble techniques to see
if even better results can be obtained.

In order to address the concerns we mentioned at the onset, we first use a content-
based recommender algorithm, thus there is no cold start issue. all of the functions
of our recommendation engine. Some of them are more diversified and precise
than others because they originate from various research departments within the
organisation. Then, the cosine similarity was introduced, which is frequently used
in industry. To improve the movie's representation for the weight of features, we
introduced TF-IIDF-DC.

47
 5.2) Future Scope

Movie genres have been included in the suggested strategy, but in the future, we
need also consider user age because movie preferences fluctuate with age. For
example, while we are young, we often like animated films over other genres.
Future versions of the proposed solution should have lower memory requirements.
Here, only different movie datasets have been used to apply the suggested
methodology. The performance can be calculated in the future and applied to the
Netflix and Film Affinity databases.

I will consider the following aspects in future work:

1. Use collaborative filtering to make recommendation

Once there is enough user data, recommendations for collaborative filtering will

be introduced. As we discussed in, collaborative filtering is dependent on user

social information, and future studies will analyze this data.

2. Include more pertinent and accurate movie features .

Common collaborative filtering recommendations replace object features with

the rating In the future, we should extract information from films that can

provide a more accurate description of the film, such as color and subtitles

3. Present the user’s dislike list of films.

User data is always useful for recommender systems. We'll keep compiling user
data and add a list of films that people don't like. We will also enter a list of films we
detest into the recommender system in order to create scores that will be added to the

48
prior result. By doing this, we can improve the recommender system's functionality.

4. Explain machine learning

Recommender systems with dynamic parameters will be studied later.

Machine learning will be used to choose the best weights and automatically
change the weights of each feature.

5. Make the recommender system a part of company.

In the future, the recommender system won't be a test-only external website.

We'll develop a programmers-only internal API. Some movie listings on the internet
will be sorted according to user reviews.

5.3) Application

Filtering and predicting only the movies that a matching user is most likely to wish
to see is the main objective of movie recommendation systems. The user
information from the system's database is used by the ML algorithms for these
recommendation systems.
Since the recommendations are particular to this person, the model doesn't require
any information about other users. This makes scaling to a huge user base simpler.
The programme may identify a user's individual preferences and offer specialized
products in which only a small percentage of other users are also interested.

According to the user's past behaviour or explicit feedback, content-based filtering

uses item features to suggest additional goods that are similar to what they already
enjoy.
hence, the applications of movie recommendation systems is to identify, filter, and
forecast the movies that a given user is most likely to find interesting. The user
information from the system's database is used by the ML algorithms for these
recommendation systems.

49
 REFERENCE

[1] Hirdesh Shivhare, Anshul Gupta and Shalki Sharma (2015), “Recommender
system using fuzzy c-means clustering and genetic algorithm based weighted
similarity measure”, IEEE International Conference on Computer, Communication
and Control.

[2] Manoj Kumar, D.K. Yadav, Ankur Singh and Vijay Kr. Gupta (2015), “A
Movie Recommender System: MOVREC”, International Journal of Computer
Applications (0975 – 8887) Volume 124 – No.3.

[3] RyuRi Kim, Ye Jeong Kwak, HyeonJeong Mo, Mucheol Kim, Seungmin
Rho,Ka Lok Man, Woon Kian Chong (2015),“Trustworthy Movie Recommender
System with Correct Assessment and Emotion Evaluation”, Proceedings of the
International MultiConference of Engineers and Computer Scientists Vol II

[4] Zan Wang, Xue Yu*, Nan Feng, Zhenhua Wang (2014), “An Improved
Collaborative Movie Recommendation System using Computational
Intelligence”,Journal of Visual Languages & Computing,Volume 25, Issue 6.

[5] Debadrita Roy, Arnab Kundu, (2013), “Design of Movie Recommendation

System by Means of Collaborative Filtering”, International Journal of Emerging
Technology and Advanced Engineering, Volume 3, Issue 4.

50