Chapter Two

Literature Review

2.1 Introduction
Among the many applications of these technologies, recommender systems have
emerged as powerful tools that help users navigate the overwhelming amount of
information available and discover personalized content tailored to their interests.
This paper delves into the realm of recommender systems, focusing specifically on
movie recommendations, and explores the latest survey and review articles that
highlight the importance of these systems in enhancing user experiences. A key
focus of this review is user satisfaction, which plays a crucial role in determining the
effectiveness of these systems. We also explore how movie recommendation
systems, designed to personalize movie suggestions for users, aim to enhance user
experience. Additionally, we discuss the importance of evaluating movie content, as
factors like genre, ratings, and user preferences significantly influence the quality
and relevance of recommendations. The review further examines various
recommendation system techniques, such as content-based filtering, collaborative
filtering, and hybrid methods, highlighting their advantages and limitations in relation
to user satisfaction. Moreover, the review addresses key challenges in the field,
including the cold-start problem, scalability, and the balance between accuracy and
diversity in recommendations, along with other factors that influence the
performance of these systems. By reviewing these studies, this research aims to
identify gaps in existing research, laying the foundation for this study, which seeks to
explore ways to improve user satisfaction in movie recommendation systems and
assess their effectiveness.

2.2 Introduction To Recommendation Systems

With the rapid growth of modern technology, the amount of data generated daily has
increased significantly, bringing us into the era of big data. While this digital
transformation has improved many aspects of life, it has also created the problem of
information overload. This occurs when people are presented with more data than
they can process, making it difficult to make decisions or find relevant information. To
tackle this challenge, methods like data mining have become essential for filtering
and organizing data effectively. One of the most practical tools in this area is the
recommendation system [ ].

A dataset is a collection of structured data used to train and test models, such as
recommendation systems. In the context of recommendation systems, datasets
typically consist of user interactions with items, such as ratings, clicks, or views,
along with additional information like user demographics or item features. These
datasets provide the foundation for analyzing patterns and making predictions about
what users might like based on past behavior.

Examples of datasets commonly used in recommendation systems include:

●​ MovieLens Dataset: A dataset of movie ratings that contains user ratings and
metadata about the movies. It is widely used for research in recommendation
systems, containing information such as movie titles, genres, and user
ratings.
●​ Amazon Product Dataset: Contains data about products purchased on
Amazon, including user reviews, ratings, and product information like
category, brand, and price. This dataset is often used to build product
recommendation systems.
●​ Netflix Dataset: A collection of user ratings for movies and TV shows on
Netflix, including metadata about the content like genres, directors, and
actors. It’s used to make personalized content recommendations for users.
●​ Goodreads Dataset: This dataset contains information about books, authors,
and user reviews. It is typically used in book recommendation systems to
suggest books based on user preferences and ratings.

These datasets help recommendation systems identify patterns in user behavior and
content preferences, enabling them to suggest relevant items.

Recommendation systems help users make decisions by suggesting items that

match their preferences, making it easier to sift through large amounts of
information. Essentially, these systems act as tools to guide users toward things they
may want or enjoy. They do this by analyzing patterns in user behavior and using
even minimal data to generate relevant suggestions. By considering factors like past
behavior, preferences, and specific features of items, recommendation systems
identify correlations to provide personalized recommendations. Today, these systems
are an integral part of platforms like Google, Amazon, Netflix, and other popular
e-commerce, music, and video services. With so many choices available, they make
it easier for users to find what they’re looking for quickly.

Challenges Of Recommendation Systems

1. Gathering User Ratings​
Many users do not provide explicit ratings, making it difficult to gauge their
satisfaction or preferences. To address this, implicit methods, such as analyzing user
behavior, are often used to predict preferences. The explicit method relies on
user-provided ratings for content. However, this approach is less effective as most
users are not inclined to rate content they interact with. Even when users provide
ratings, the number of participants is often too limited, resulting in a sparse
User-content rating table that complicates decision-making. Implicit methods infer
ratings from user behavior, such as viewing or interaction patterns. Many
recommendation systems combine both explicit and implicit data for improved
predictions (Felfemig, Friedrich, Jannach, & Zanker, 2011; Salter & Antonopoulos,
2006). Despite these efforts, challenges remain, highlighting the need for further
innovation to enhance the precision and effectiveness of these approaches [ ].

2. Cold Start Problem​

This problem occurs in scenarios where little to no prior information is available:

●​ New User Problem: First-time users lack a history, making personalized

recommendations difficult.
●​ New Item Problem: Newly added options with no interaction history are
challenging to recommend.
●​ New System Problem: When a system is launched without any prior user or
content data, recommendations are ineffective. Addressing this requires using
external data or employing creative solutions like hybrid approaches [ ].

3. Sparsity Problem​
A sparsity issue arises when the user-content interaction data contains insufficient
information, often due to users failing to rate or interact with much content. Sparse
data affects the accuracy of recommendations and highlights the need for reliable
algorithms that can work well with limited data [ ].

4. Lack of Data​
Recommendation systems rely on rich datasets to perform well. Smaller or newer
platforms without access to extensive user or content data struggle to provide
relevant suggestions, making data collection and enhancement essential for better
performance [ ].

5. Changing Data and Preferences​

Users’ preferences and behaviors evolve, and systems must adapt dynamically to
these changes. Failure to respond to shifting trends or interests, such as life stage
changes or changes in external factors like holidays, cultural events, or economic
conditions, can make recommendations irrelevant [ ].

6. Unpredictable contents​
When new or unusual content is introduced without enough past data, the
recommendation system has a hard time suggesting it. This happens because the
system relies on patterns from past interactions to make suggestions. For instance, if
a new product or movie has no ratings or user engagement yet, the system won't
know how to connect it to what a user might like. As a result, these new products or
movies might not get recommended as much because there isn’t enough information
to make an accurate prediction about them [ ].

7.Scalability Issues​
Handling a growing number of users and items while maintaining system
performance is a critical challenge. Algorithms that work well on smaller datasets
may struggle with larger ones. Optimizing the structure of both hardware and
software is essential to ensure efficiency as the system expands [ ].
2.3 Movie Recommendation Systems
When it comes to movies, recommendation systems work by suggesting films based
on a user’s preferences and past viewing history. User profiles in movie
recommendation systems are created by collecting details like age, gender, location,
and personal preferences. This information helps the system understand what kind
of movies someone might enjoy. Similarly, movie profiles are built using features like
genre, director, cast, release year, and language. For example, a movie might be
tagged as "Action," "Directed by Christopher Nolan," or "Starring Leonardo
DiCaprio." How these profiles are created is crucial because it directly impacts how
accurate and effective the recommendations are. By matching the traits in a user’s
profile with the features of different movies, the system can suggest films that better
fit the user’s tastes [ ]. For instance, a child would likely receive recommendations for
cartoons and animations, which are more popular with younger viewers. The system
can also adjust its suggestions by considering what other children of the same age
are watching, ensuring that the recommendations are relevant.

In addition, movie recommendation systems address the problem of information

overload by filtering out unnecessary content and highlighting the most relevant
options. By analyzing a user’s profile and past behavior, the system can predict
which movies are likely to be of interest, helping users make quicker and more
informed choices.This personalized approach not only improves the decision-making
process but also enhances user satisfaction, making recommendation systems an
essential feature of modern streaming platforms.

They also help users discover movies that align with their unique preferences, saving
time and effort spent searching for something to watch. Offering personalized
suggestions that cater to individual tastes ensures users feel more connected to the
platform, creating a smoother and more enjoyable experience.

By presenting relevant and appealing content, these systems encourage users to

spend more time on the platform. Suggested options reduce decision fatigue, making
it easier for users to quickly find something they want to watch, which in turn
increases engagement and keeps them coming back for more.

For streaming platforms or cinemas, these systems can drive key business
outcomes. By increasing movie viewership, subscriptions, or purchases through
targeted recommendations, they directly contribute to revenue growth. Moreover,
they help improve customer retention by ensuring users stay satisfied and engaged
with content that appeals to them.

They also introduce users to new genres, directors, or movies they may not have
discovered on their own. By balancing popular content with niche films, these
systems broaden the range of choices, allowing users to explore different types of
content they might not have considered before.
These systems help streaming platforms maximize the value of their entire movie
library, recommending less popular or older content that users might otherwise
overlook. By surfacing hidden gems, they help reduce the "long tail" problem,
ensuring that all content, both mainstream and obscure, gets the attention it
deserves.

2.4 Techniques Of Recommendation Systems

Movie recommendation systems work by filtering out irrelevant information and
focusing on data that aligns with a user’s preferences or content features [ ]. They
analyze patterns in user behavior and the characteristics of movies to identify what
someone might enjoy. This process involves gathering details like viewing history,
ratings, user demographics, and sometimes even factors like the time of day or the
device being used. By narrowing down the options, these systems make it easier for
users to discover movies that match their interests.

To achieve this, recommendation systems use three main techniques: Collaborative

Filtering, Content-Based Filtering, and Hybrid Filtering. Collaborative Filtering finds
patterns in user behavior, such as identifying users with similar tastes and
recommending what they enjoy. Content-Based Filtering looks at the features of
movies like genre, cast, or director and suggests similar content based on what a
user has liked before. Hybrid Filtering combines both methods to improve
recommendations and overcome the weaknesses of using only one method.These
techniques create a more personalized experience, helping users save time and
discover content they might not have found otherwise. Understanding how each of
these methods works helps explain the ideas behind today’s recommendation
systems.

Collaborative Filtering

Collaborative filtering uses historical user activity to predict movies a user might
enjoy. It analyzes user information, such as the movies they’ve watched, searched
for, or rated, and compares it with the activities of other users with similar
preferences. For example, in movie recommendation systems, demographic details
like age, gender, or ethnicity are combined with past viewing and search history to
suggest movies that align with the interests of similar users. By identifying patterns
among users with shared tastes, it predicts recommendations for the target user.

There can be countless users in the system. This technique identifies users with
similar preferences by analyzing the ratings they have given to specific items. By
comparing these ratings, the system finds patterns of similarity between users. The
strategy relies on the ratings provided by users across a broad catalog of items. This
catalog, often referred to as the user-item matrix, serves as the foundation for
generating relevant recommendations. In the context of movie recommendation
systems, this catalog could be referred to as the user-movie matrix.
Collaborative filtering was first introduced in 1991 by Goldberg et al. with the
development of the Tapestry system. Tapestry was an early attempt at creating a
collaborative recommendation system, designed for smaller user groups. However, it
had its limitations, requiring a lot of user input and not being very scalable for larger
groups. While Tapestry helped to demonstrate the potential of recommendation
systems, it also highlighted the need for improvements.

Later, systems like Grouplens advanced the concept by using scoring-based

collaborative filtering, which proved to be much more effective. Grouplens and
similar systems were capable of recommending content like news and movies by
analyzing users' preferences and identifying patterns in their behavior. Today,
collaborative filtering is widely used in e-commerce platforms like Amazon and
CDNow, helping users find products and content that match their interests.

In the context of movies, it has become one of the most popular methods for
recommending films on streaming services, offering personalized suggestions to
enhance the user experience. For example, platforms like Netflix and IMDb rely on
collaborative filtering techniques to suggest movies, helping users find films they
might enjoy based on the preferences of other users with similar tastes.

The power of collaborative filtering lies in its ability to sift through large amounts of
data and provide personalized recommendations, making it easier for users to
discover things they'll enjoy without needing to search through endless options. It's
become a standard technique for recommending everything from movies to products
across online platforms.

Collaborative filtering is considered one of the most effective recommendation types

in recommendation systems and is widely employed. However, it has its limitations.
One of the main challenges is the cold start problem, where the system struggles to
make recommendations if a user hasn’t provided enough information or if there is too
little data about an item.

Another issue is that the system’s accuracy can be limited, as people with similar
demographic profiles don’t always have similar tastes. Moreover, because
collaborative filtering relies on finding similarities between users or products, it can
sometimes lead to repetitive suggestions, reducing the variety of recommendations
offered.

Researchers have worked on overcoming the challenges of collaborative filtering,

like data sparsity and the cold start problem. Through various studies, they’ve
introduced new methods to improve accuracy, scalability, and diversity in
recommendations, making suggestions more relevant and personalized for users.
Mehrbakhsh Nilashi proposed a recommendation method that combines
collaborative filtering with ontology and dimensionality reduction techniques. This
approach helps address the sparsity and scalability issues in collaborative filtering.
Their experiments showed improvements in predictive accuracy and the
performance of movie recommendation systems [ ].

Miguel G. Silva et al. developed a collaborative filtering method that groups users by
analyzing patterns in how they rate items. For example, if two users consistently rate
similar items in the same way, they are grouped together. This grouping helps predict
preferences for users even when data is sparse or inconsistent, leading to more
accurate and personalized recommendations [ ].

Daniel A. Galron et al. introduced a deep learning method that works by improving
how the system identifies similarities between users or items. Instead of relying on
traditional metrics, their approach uses advanced neural networks to process the
available data more effectively. This helps overcome challenges in datasets where
user activity is limited or spread out, resulting in better recommendations [ ].

Ali Fallahi RahmatAbadi and Javad Mohammadzadeh explored how deep learning
can solve common collaborative filtering problems. For instance, they highlighted
how neural networks can predict preferences for new users or items (cold start) by
analyzing related data, such as item features or user profiles. Additionally, they
showed how these methods can handle and process large amounts of data more
efficiently, improving scalability for bigger systems [ ].

Dongsheng Li et al. focused on making collaborative filtering algorithms more stable

and reliable. They proposed a method that ensures the system processes data
consistently, reducing errors caused by changes or fluctuations in the input. This
makes recommendations more accurate and dependable, even when working with
large datasets or rapidly changing user data [ ].

Collaborative filtering plays an important role in recommendation systems and

continues to evolve to keep up with the increasing complexity of user data. Its
strength lies in its ability to find meaningful patterns, making it a powerful tool for
creating personalized experiences across different platforms.

Content-Based Filtering

Content-Based Filtering is also known as cognitive filtering. This filtering

recommends contents to the user based on his past experience. For example, if a
user likes only action movies then the system recommends him only action movies
similar to it which he has highly rated. The broader explanation could be, suppose
the user likes only politics related content so the system suggests the websites,
blogs or the news similar to that content. A content based filtering technique works
by making movie proposals to the user based on the content in the movies. It
recognizes that clustering in the collaborative filtering recommendations may not
match the preferences of the users [ ]. This technique takes advantage of similarities
between movies and assumes that if a viewer enjoys one movie, they are likely to
appreciate comparable movies sharing attributes like themes, actors, or styles. Even
if no explicit genre preference is saved in the user’s profile (a collection of their past
interactions, ratings, and preferences) the system can infer these preferences from
available data, such as viewing history or ratings [ ].

Unlike collaborative filtering, content based filtering does not face new user
problems. It does not have other user interaction in it. It only deals with a particular
user’s interest. Collaborative filtering, on the other hand, assumes that grouping
users with similar demographic characteristics will result in effective
recommendations. However, it recognizes that grouping in collaborative filtering,
recommendations may not match the preferences of the users [7].For example, The
tastes and preferences of people with similar demographic characteristics can vary
significantly, what person X likes may not align with what person Y enjoys watching [
].Content based filtering first checks the user preference and then suggests him with
the movies or any other product to him. It only focuses on a single user’s ideas,
thoughts and gives suggestions based on his interest. So if we talk about movies,
then the content based filtering technique checks the rating given by the user. The
approach checks which movies are given high ratings by the user by checking the
genre categories in the user profile. After analysing the user profile, the technique
recommends movies to the user according to his taste [ ]. The fundamental
principles of content-based filtering can be broken down into two key steps: (1)
analyzing the movies a specific user enjoys, identifying common attributes such as
genre, director, cast, or themes, and storing these preferences in the user’s profile;
and (2) comparing the attributes of other movies with the user’s profile to
recommend films that closely match their preferences [ ].

However, this type of recommendation approach carries the risk of excessive

personalization, meaning that the recommendations received by users are limited to
products that are highly similar to their user profiles [ ]. This is generally referred to
as the overspecialization problem, as users are rarely recommended anything
outside their profiles. It can also become difficult for the system to identify and
suggest specific features, such as images or movies from certain genres, especially
if those genres are not explicitly defined in the user’s profile. Without sufficient
context, the system may struggle to provide a diverse range of recommendations,
limiting the variety of options presented to the user.. Additionally, It is easy to miss
recommending an item to a user as there is not enough information about that item
[ ].
To overcome the overspecialization problem in content-based filtering, researchers
have introduced innovative solutions. These efforts, highlighted in various studies,
aim to make recommendations more diverse and meaningful, giving users a broader
range of options to choose from.

Jieun Son and Seoung Bum Kim suggested a way to improve content-based filtering
for movie recommendation systems by using multiattribute networks. These
networks include detailed information about the movies being recommended. By
analyzing all the attributes through network analysis, their method recommends a
wider variety of movies, effectively solving the overspecialization problem. Their
results also showed improvements in dealing with issues like sparsity and scalability
compared to traditional content-based filtering methods. By leveraging movie
attributes like cast, keywords, crew, and genres, they aim to enhance the
movie-watching experience for users, saving them time and effort in searching for
movies that align with their tastes [ ].

Similarly, in the study “Providing Entertainment by Content-based Filtering and

Semantic Reasoning in Intelligent Recommender Systems” by Yolanda
Blanco-Fernandez and her team tackled the overspecialization problem by
incorporating semantic reasoning into content-based filtering. Their approach used
additional contextual and meaningful information about movies, such as relationships
between genres, themes, or user preferences, to provide more diverse and relevant
recommendations. This ensured that users were not limited to overly specific
suggestions, enhancing the variety and appeal of the recommendations [ ]. Previous
studies on movie recommendation systems have mainly relied on collaborative
filtering, which might face limitations when dealing with new or niche movies with
sparse user ratings. Additionally, content based filtering approaches have shown
promise in providing personalized recommendations but might overlook the social
aspects of user preferences. Previous studies on movie recommendation systems
have mainly relied on collaborative filtering, which might face limitations when
dealing with new or niche movies with sparse user ratings. Additionally, content
based filtering approaches have shown promise in providing personalized
recommendations but might overlook the social aspects of user preferences.
Therefore, there is a need to address these challenges and create a
recommendation system that combines the strengths of both approaches. [ ].

Hybrid Filtering
This filtering is an information filtering system that takes ratings of the movies as
input from the users and then applies the collaborative and content based filtering
and generates recommendation lists [49]. It is a combination of the two techniques
i.e. collaborative filtering and content based filtering. It is superior because it
achieves higher performance in making the suggestions and also a faster
computational time [ ]. When only the single method i.e. the collaborative filtering or
content based filtering alone cannot solve the problem then hybrid filtering concept
comes into picture. By using hybrid filtering many problems of collaborative filtering
and content based filtering can be resolved. For example, the cold start problem in
collaborative filtering and the lack of user preference information in content-based
filtering are significant challenges. So if we apply content based filtering and then
use collaborative filtering it can be a solution to it. So making it hybrid can resolve
the problem.

Yang et al. put forth a hybrid approach based on social similarity and item attribute.
The author used collaborative filtering methods combined with social similarities and
genres of the movie. They used a model called BPR-MF (Bayesian Personalized
Ranking - Matrix Factorization) to address the problem of sparse data. BPR-MF
(Bayesian Personalized Ranking Matrix Factorization) is a recommendation model
that predicts user preferences by ranking items rather than predicting ratings. It
breaks down a user-item interaction table into two smaller tables, optimizing for item
rankings based on user interactions. It's particularly effective with sparse data and
implicit feedback. The proposed method works in two stages. First the BPR-MF
model is used to obtain the candidate set which refers to a group of potential items
that could be recommended to a user, based on the ratings from the training dataset.
After finding the candidate set, the unknown ratings are predicted using the existing
ratings. Then the ratings are sorted and the final candidate set for each user is
obtained. Each set has several top items. In the second stage the movies are
recommended to users using the feature selection TF-IDF method. TF-IDF is a
method used to evaluate how important features (like words or attributes) are in a
dataset, based on how often it appears and how rare it is across all datasets. It helps
find similarities between items or users in recommendation systems.. The result
shows that using BPR-MF shows more accurate results rather than collaborating
filtering.

Priscila Valdiviezo and J. Bobadilla proposed a method that combines various user
ratings and demographic information like age, gender and occupation and combined
them into one matrix model. Then collaborative filtering is used to find out the
missing ratings. The main idea used here is to improve the overall rating prediction.
Here MAE (Mean Absolute Error) is used to measure the performance of the
proposed approach. The data sparsity problem is also solved by using demographic
features of the user and the item.

F.Deng et al. defined an approach to calculate user’s potential preference based on

hybrid features like user generated features, image visual features and transforming
user item rating into hybrid feature ratings [ ]. The experiments performed here
shows that proposed work gives better results on sparse dataset and has higher
efficiency on large dataset.

The hybrid recommendation system combines the strengths of various techniques,

offering a more robust and accurate approach to personalized recommendations. By
leveraging multiple methods, it can address the limitations of each individual
technique. However, while hybrid models are powerful, they come with challenges.
They often require significant computational resources and may involve complex
processing, leading to potential performance issues. Additionally, combining different
methods can sometimes yield conflicting results, requiring careful tuning and
balancing to ensure the system provides consistent and reliable recommendations.

2.5 User Satisfaction in Recommendation Systems

This is the era of automation and artificial intelligence, where users rely on
automated agents for movie, song, and product recommendations. These
personalized systems are increasingly shaping daily life. According to previous
studies, many end-users struggle to understand how these 'black box' systems work.
The outputs of recommendation systems (such as movie suggestions) are often not
explainable, and the underlying algorithms remain incomprehensible to users.
Additionally, as these systems are adaptive, there is no standardized metric to
assess their personalized performance. Therefore, evaluating user satisfaction
becomes the ultimate method for measuring system effectiveness.

User satisfaction is also commonly used to evaluate the success of information

systems. According to Zipf’s principle of least effort, a fundamental principle of
human action is to exert the least effort to do things. The opposite of this principle is
information overload—where users are bombarded with more information than they
can process in a given timeframe. Personalized recommendation systems help
mitigate information overload by offering users information aligned with their
preferences. In other words, by adhering to the principle of least effort, these
systems reduce cognitive load and enhance user satisfaction.

User satisfaction is achieved when the user’s goals are aligned with the system’s
suggestions. Understanding user perception is key to building a quality
recommender system. Providing the right recommendations at the right time can
boost user satisfaction and encourage engagement. From a reliability standpoint,
previous studies have shown that when novice users understand the reasoning
behind the system's suggestions and can anticipate outcomes, along with having
some control over the system, it helps build trust. This, in turn, enables users to
personalize the agent more effectively. This section will explore the factors
influencing user satisfaction in recommendation systems and the methods used to
measure it effectively.