INTRODUCTION TO

ARTIFICIAL INTELLIGENCE
TOPIC: MOVIE RECOMMENDATIONS SYSTEM

GROUP 3
Team Members
Vu Thuong Tin 20230091 Chu Anh Duc 20230081
Tran Quang Hung 20235502 Phan Dinh Trung 20230093

Instructor: Assoc. Prof. Le Thanh Huong

Major: Data Science and Artificial Intelligence

School: School of Information and Communications Technology

HANOI - 11 - 2024
 TABLE OF CONTENTS

CHAPTER 1. INTRODUCTION............................................................. 2

1.1 Problem Statement................................................................................ 2

1.2 Background and Problems of Research................................................... 2

1.3 Research Objectives and Conceptual Framework..................................... 3

CHAPTER 2. LITERATURE REVIEW .................................................. 6

2.1 Scope of Research ................................................................................ 6

2.2 Related Works ...................................................................................... 7

CHAPTER 3. BASIC MATHEMATICS .................................................. 8

3.1 Content - Based Recommendations ........................................................ 8

3.2 Neighborhood - Collaborative Filtering .................................................. 9

3.3 Matrix Factorization ............................................................................. 11

CHAPTER 4. OVERVIEW ..................................................................... 13

4.1 Abstract ............................................................................................... 13

4.2 Data Processing.................................................................................... 13

4.3 Model Training - Algorithms ................................................................ 15

CHAPTER 5. NUMERICAL RESULTS.................................................. 20

5.1 Evaluation Parameters........................................................................... 20

5.2 Result of Evaluating.............................................................................. 21

CHAPTER 6. CONCLUSIONS ............................................................... 24

6.1 Summary ............................................................................................. 24

6.2 Suggestion for Future Works ................................................................. 24

 CHAPTER 1. INTRODUCTION

1.1 Problem Statement

Recommendation systems have become an integral part of modern digital platforms,
offering personalized suggestions to users based on their preferences and behavior.
In the domain of movies, these systems play a crucial role in enhancing user
experience by helping individuals navigate the overwhelming number of options
available on streaming services, online stores, and other media platform. By leveraging
techniques such as collaborative filtering, content-based filtering, and hybrid approaches,
movie recommendation systems analyze vast datasets of user interactions, movie
metadata, and contextual information to deliver tailored recommendations. The
primary objective is to bridge the gap between user preferences and available
content, fostering engagement, satisfaction, and retention.

Figure 1: Polular Recomendations System

1.2 Background and Problems of Research
Otherwise, recommendation systems is a fairly broad area of Machine Learning
and is younger than Classification because the internet has only really exploded in
the last 10 - 15 years. There are two main entities in the Recommendation Systems
users, and items or users are users. Items are products, such as movies, songs,
books, clips, or other users in the friend suggestion problem. The main purpose
of Recommender Systems is to predict a user’s interest in a certain item, thereby
having a suitable recommendation strategy.
There are various challenges faced by Recommendation System. These challenges
are Cold Start problem, Data Sparsity, Scalability.
Cold Start Problem: It needs enough users in the system to find a match. For
instance, if we want to find similar user or similar item, we match them with the set
of available users or items. At initial stage for a new user, his profile is empty as he
has not rated any item and the system do not know about his taste, so it becomes

2
CHAPTER 1. INTRODUCTION

difficult for a system to provide him recommendation about any item. Same case
can be with new item, as it is not rated by any user because it’s new for the user.
Both these problem can be resolved by implementing hybrid techniques.
Data Sparsity: The user or rating matrix is very sparse. It is very hard to
find users that have rated the same items because most of the user does not rate
the items. So it becomes hard to find set of users who rate the items. To give
recommendation is really tough when there is less information about any user .
Scalability: Collaborative Filtering use massive amount of data to make reliable
better which require more number of resources. As information grows exponentially
processing becomes expensive and inaccurate result from this Big data challenge.

Figure 2: Structure of Recommendations system

1.3 Research Objectives and Conceptual Framework
There are a number of datasets that are available for recommendation research.
Among of them, the MovieLens dataset is probably one of the more popular ones.
MovieLens is a non-commercial web-based movie recommender system. It is
created in 1997 and run by GroupLens, a research lab at the University of Minnesota,
in order to gather movie rating data for research purposes. MovieLens data has
been critical for several research studies including personalized recommendation
and social psychology.

3
 CHAPTER 1. INTRODUCTION

Figure 3: Comparision of ML 100k and ML 1M

When working with datasets like ML 100K and ML 1M, there are a few notable
challenges to be aware of. First, the large size of ML 1M requires more powerful
computational resources and longer processing times, while ML 100K can suffer
from overfitting due to small and lack of diverse data. Second, data sparsity reduces
the effectiveness of matrix-based algorithms and makes it difficult to predict for
less interactive movies or users, especially the cold-start problem when dealing
with new users or movies. Third, the uneven quality of the data, with subjective
and inconsistent ratings, makes it difficult to distinguish between users who dislike
a movie and those who have never watched it. Finally, balancing accuracy and
performance is a major challenge, especially when large datasets require more
processing complexity but offer the potential to improve recommendation performance.
To overcome this, methods such as data size reduction, using association models,
and exploiting additional information need to be applied to improve performance
and accuracy.
We have proposed the following algorithms Content-Based Recommendations,
Neighborhood - Based Collaborative Filtering, Matrix Factorization.
Content - Based Recommendations: The Content-Based Recommendation
algorithm focuses on analyzing the characteristics of movies and the user’s interaction
history to provide suggestions. This algorithm emphasizes finding movies similar
to those the user has liked or watched before by analyzing their content. There are
3 steps.
1. Analyzing movie features: Each movie is described by a set of attributes
(features), such as genres, director, main actors, release year, Plot descriptions
or related keywords. These attributes are represented as content vectors for
each movie.

4
CHAPTER 1. INTRODUCTION

2. User profile: A user profile is built based on their watch history, ratings, or
interactions. The user profile is represented by aggregating the content vectors
of the movies they have rated highly.
3. Finding similar movies: After creating the user profile, the algorithm searches
for movies with content vectors similar to the user profile vector (usually using
similarity measures like cosine similarity). Movies with high similarity scores
are recommended.
Neighborhood - Based Collaborative Filtering: Neighborhood-Based Collaborative
Filtering is a widely used technique in recommendation systems that relies on the
concept of similarity between users or items (movies) to make predictions. Unlike
Content-Based Recommendation, which focuses on the attributes of movies, Collaborative
Filtering works by analyzing user-item interactions (e.g., ratings) to identify patterns
of shared preferences. There are 2 steps.
1. User-Based Collaborative Filtering: Focuses on identifying similar users
based on their past interactions, key idea - Users who have rated or liked
similar movies in the past are likely to have similar preferences in the future.
Compute similarity between the target user and other users (e.g., using cosine
similarity or Pearson correlation on their rating vectors), identify a set of
neighboring users. Then, we aggregate the ratings of the neighboring users
to predict the target user’s rating for a movie.
2. Item-Based Collaborative Filtering: Focuses on finding similar movies based
on user ratings, key idea - Movies that are rated similarly by multiple users are
likely to have similar appeal. Compute similarity between movies, for a given
movie, find a set of neighboring movies. Then, we predict the user’s rating for
a movie based on their ratings of similar movies.
Matrix Factorization: Matrix Factorization is a powerful and widely used
technique in recommendation systems that leverages linear algebra to predict missing
entries in a user-item interaction matrix (e.g., a matrix of user ratings for movies).
It is particularly effective for large, sparse datasets and forms the basis of modern
recommendation algorithms.

5
 CHAPTER 2. LITERATURE REVIEW

2.1 Scope of Research

The MovieLens 100k dataset is a cornerstone in recommender systems research,
providing a rich collection of 100,000 user ratings across 1,682 movies from 943
unique users. Its structured yet compact size offers a practical platform for experimenting
with and comparing various recommendation algorithms. This research aims to
explore the application of Content-based recommendations, Collaborative filtering,
and Matrix factorization methods to maximize accuracy, scalability, and user
satisfaction within the MovieLens 100k dataset’s context.
Research Objectives
1. Understand Dataset Properties
+ Analyze user and movie distributions, rating patterns, and sparsity levels.
+ Evaluate potential biases, such as popularity bias or rating concentration on
specific movies or users.
2. Algorithm Exploration
+ Content-based recommendations: Leverage movie metadata (e.g., genres,
titles) to recommend movies similar to those a user has previously rated.
+ Collaborative filtering:
- User - user and item - item methods to identify similarities in rating patterns.
- Address the cold-start problem by exploring hybrid approaches.
+ Matrix factorization: Use techniques like Singular Value Decomposition
(SVD) and Alternating Least Squares (ALS) to uncover latent factors representing
user and movie characteristics.
3. Model Implementation and Optimization
+ Implement selected algorithms using frameworks like Scikit-learn, Surprise,
and TensorFlow.
+ Optimize hyperparameters to enhance recommendation quality, addressing
overfitting and underfitting.
4. Evaluation Metrics and Benchmarking
+ Use metrics such as Mean Absolute Error (MAE), Root Mean Squared
Error (RMSE), and Precision@K for quantitative performance evaluation.
+ Perform comparative analysis to assess trade-offs between computational

6
CHAPTER 2. LITERATURE REVIEW

efficiency and recommendation accuracy.

5. Practical Deployment Considerations
+ Evaluate the feasibility of real-world deployment by analyzing algorithm
scalability and responsiveness to dynamic datasets.
+ Explore opportunities for integrating contextual information to improve
personalization.
2.2 Related Works
Recommendation systems have become a cornerstone in personalized content
delivery, with applications spanning various domains such as e-commerce, online
streaming platforms, and social networks. In this section, we review existing
applications, focusing on Netflix’s recommendation system as a benchmark due
to its prominence and sophistication. We also discuss other relevant studies in
recommendation algorithms to contextualize the research scope.
1. Netflix Recommendation System Netflix’s recommendation system is a leading
example of how advanced algorithms can drive user engagement and business
success. Its multi-layered architecture integrates various recommendation
approaches
2. Other Applications and Studies Several other platforms and studies contribute
valuable insights into recommendation system research
+ Amazon and E-Commerce Recommendations: Amazon’s system heavily
relies on item-item collaborative filtering, prioritizing scalability for its vast
catalog. The use of implicit feedback and co-purchase data has proven effective
in generating real-time suggestions.
+ YouTube and Video Recommendations: YouTube employs a two-stage
architecture: candidate generation using collaborative filtering and ranking
based on deep learning models. The system factors in watch time, clicks, and
engagement metrics to improve relevance.
+ Spotify and Music Recommendations: Spotify integrates collaborative
filtering, audio signal processing, and natural language processing to create
personalized playlists and suggestions. Their system exemplifies how domain-
specific features (e.g., acoustic properties) can enhance user satisfaction.

7
 CHAPTER 3. BASIC MATHEMATICS

3.1 Content - Based Recommendations

1. Ridge regression
Loss function
The loss function of the Ridge regression is defined as:

N
1 X
2 λ
L(w) = yi − x ⊤
i w + ∥w∥22
2N 2
i=1

Explanation
Least squares error
N
1 X
2
yi − x ⊤
i w
2N
i=1

+ This is the sum of squared errors between the actual value yi and the
predicted value x⊤
i w.

+ This term measures the accuracy of the model.

ℓ2 Regularization term
d
λ λX 2
∥w∥22 = wj
2 2
j=1

+ This term penalizes large values of the weights w, helping to reduce

overfitting.
+ Large λ: prioritizes reducing model complexity.
+ Small λ: focuses on minimizing prediction error.
Objective The goal is to optimize w to minimize the loss function L(w):

w∗ = argmin L(w)
w

Closed-Form solution: Given the feature matrix X (N × d) and the target

vector y (N × 1), the solution to Ridge Regression is:

−1
w∗ = X⊤ X + λI X⊤ y

where:

8
CHAPTER 3. BASIC MATHEMATICS

+ I: identity matrix.
+ λ: regularization coefficient.
2. Utility matrix
Formally, a utility matrix Y is a matrix of size M × N , where:
+ M : Number of items.
+ N : Number of users.
+ ymn : The utility (e.g., rating, score, or preference) assigned by user n to
item m.
Key keatures of a utility matrix
(a) Sparse representation
In most real-world applications, the utility matrix is sparse because:
+ Not all users rate or interact with all items.
+ Many entries ymn are unknown or missing.
(b) Numerical entries
The entries in the matrix are typically numerical, such as:
+ Ratings (e.g., 1–5 stars in a movie-rating system).
+ Binary values (e.g., 0 for no interaction, 1 for interaction).
+ Implicit scores (e.g., frequency of interaction or purchase history).
(c) Approximation:
The goal of many machine learning or optimization methods (e.g., matrix
factorization) is to approximate the missing values in the utility matrix by
learning patterns from the observed entries.
3.2 Neighborhood - Collaborative Filtering
1. Similarity function Is an important concept in machine learning, data processing,
and problems involving the comparison of two objects. This function measures
the degree of similarity between two objects (e.g., vectors, data points, or sets)
based on specific criteria.
2. General formula A similarity function sim(a, b) takes two objects a and b as
input and returns a value within a defined range, typically:

sim(a, b) ∈ [0, 1] or sim(a, b) ∈ [−1, 1]

9
 CHAPTER 3. BASIC MATHEMATICS

Cosine similarity: Measures the angle between two vectors in a feature space.
It is calculated by:
a·b
simcosine (a, b) =
∥a∥∥b∥
The value ranges from [−1, 1].
3. Prediction rating
Formula
The general formula for the predicted rating r̂ui in a recommendation system
is:

r̂ui = f (u, i)

Where:
+ r̂ui is the predicted rating for user u on item i.
+ f (u, i) is a function that calculates the predicted rating based on user u’s
preferences and item i’s characteristics.
In the context of different methods, the formula can vary :
User-User Collaborative filtering

P
v∈N (u) sim(u, v) · rvi
r̂ui = P
v∈N (u) |sim(u, v)|

Where:
+ N (u) is the set of users most similar to user u.
+ sim(u, v) is the similarity between the user u and the user v .
+ rvi is the rating that user v gave to item i.
Item-Item Collaborative filtering

P
j∈N (i) sim(i, j) · ruj
r̂ui = P
j∈N (i) |sim(i, j)|

Where:
+ N (i) is the set of items most similar to item i.
+ sim(i, j) is the similarity between item i and item j .
+ ruj is the rating that user u gave to item j .

10
CHAPTER 3. BASIC MATHEMATICS

3.3 Matrix Factorization

1. Gradient’s Descent: Given the loss function for matrix factorization, the
updates for the user and item feature matrices P and Q using gradient descent
are as follows:
For the User feature Matrix P
The gradient of the loss function with respect to pik is

∂ Loss X
= −2 (rij − pTi qj )qjk + 2λpik
∂pik
j∈κi

Where:
+ κi is the set of items rated by user i
+ rij is the actual rating for user i and item j
+ pTi qj is the predicted rating for user i and item j
+ λ is the regularization parameter
+ qjk is the feature of item j corresponding to the latent factor k
The update rule for pik is:

∂ Loss
pik := pik − η
∂pik

where η is the learning rate.

For the Item feature Matrix Q
The gradient of the loss function with respect to qjk is:

∂ Loss X
= −2 (rij − pTi qj )pik + 2λqjk
∂qjk
i∈κj

Where:
+ κj is the set of users who rated item j
+ pTi qj is the predicted rating for user i and item j
+ pik is the feature of user i corresponding to the latent factor k
+ λ is the regularization parameter
The update rule for qjk is:

11
 CHAPTER 3. BASIC MATHEMATICS

∂ Loss
qjk := qjk − η
∂qjk

where η is the learning rate.

2. Singular Value Decomposition - SVD

r̂ui = pTu qi

Where:
+ pu is the latent feature vector for the user u
+ qi is the latent feature vector for item i
+ pTu qi is the dot product between the vectors of features of the user and
the item

12
 CHAPTER 4. OVERVIEW

4.1 Abstract
This study outlines a comprehensive methodology pipeline for building and
deploying a recommendation system using the MovieLens 100k dataset. The
pipeline is divided into four major stages: Data Processing, Model Training, Deployment.
The recommendation algorithms used in this research are Content-based Filtering,
Collaborative Filtering, and Matrix Factorization. Each stage plays a crucial role
in ensuring the effectiveness, accuracy, and scalability of the recommendation
system. Below is a detailed description of each step in the methodology.
4.2 Data Processing
We will use dataset Movielen100k: https://files.grouplens.org/
datasets/movielens/ml-100k.zip
1. Collection Data
Here are brief descriptions of the data.
• u.data – The full u data set, 100000 ratings by 943 users on 1682
items. Each user has rated at least 20 movies. Users and items are
numbered consecutively from 1. The data is randomly ordered. This
is a tab-separated list of:
user id|item id|rating|timestamp

• u.info – The number of users, items, and ratings in the u data set.
• u.item – Information about the items (movies); this is a tab-separated
list of:
movie id|movie title|release date|video release date|
IMDb URL|unknown|Action|Adventure|Animation|
Children’s|Comedy|Crime|Documentary|Drama|Fantasy|
Film-Noir|Horror|Musical|Mystery|Romance|Sci-Fi|
Thriller|War|Western

The last 19 fields are the genres, a 1 indicates the movie is of that
genre, a 0 indicates it is not; movies can be in several genres at once.
The movie ids are the ones used in the u.data file.

13
 CHAPTER 4. OVERVIEW

• u.genre – A list of the genres.

2. File Descriptions
• u.user – Demographic information about the users; this is a tab-
separated list of user id | age | gender | occupation
| zip code. The user IDs are the ones used in the u.data data
set.
• u.occupation – A list of the occupations.
• u1.base, u1.test, ..., u5.base, u5.test – The data sets u1.base
and u1.test through u5.base and u5.test are 80%/20% splits
of the u.data into training and test data. Each of u1, ..., u5 have
disjoint test sets; this is for 5-fold cross-validation (where you repeat
your experiment with each training and test set and average the results).
These data sets can be generated from u.data by mku.sh.
• ua.base, ua.test, ub.base, ub.test– The data sets ua.base,
ua.test, ub.base, and ub.test split the u.data into a training
set and a test set with exactly 10 ratings per user in the test set. The
sets ua.test and ub.test are disjoint. These data sets can be
generated from u.data by mku.sh.
• allbut.pl – The script that generates training and test sets where
all but n of a user’s ratings are in the training data.
3. Splitting Data

Figure 4: Split data

The dataset used in this experiment is the MovieLens 100K dataset, which
includes user ratings for movies. It was split into two parts: the training
data (ua.base) containing 90,570 ratings, and the test data (ua.test) containing

14
CHAPTER 4. OVERVIEW

9,430 ratings. The training data is used to build and train the model, while
the test data is used to evaluate its performance on unseen ratings. The data
was read using Python’s pandas library with tab-separated values, and then
converted into NumPy arrays for efficient computation. This split ensures that
the model is tested for generalization and not just for memorization of the
training data, providing a reliable estimate of its performance on new, unseen
data.
4.3 Model Training - Algorithms
There are used to 3 Algorithms
1. Content - Based Recommendations
Problem: Suppose the number of users is N , the number of items is M , and
the utility matrix is represented by matrix Y. The entry in row m, column
n of Y is the level of interest (e.g., rating) of user n for item m, which the
system has collected. The matrix Y is sparse and contains many missing
entries corresponding to values the system needs to predict.
Additionally, let R be the rated or not matrix, which indicates whether a user
has rated an item or not. Specifically, rij equals 1 if item i has been rated by
user j , and equals 0 otherwise.
Linear Model
Suppose we can find a model for each user, represented by a column vector
of coefficients wn and a bias term bn , such that the level of interest of a user
for an item can be calculated by a linear function:

ymn = xm wn + bn (4.1)

(Note that xm is a row vector and wn is a column vector.)

For a given user n, let the training set be the set of observed entries in yn . We
can construct the loss function similar to Ridge Regression as follows:

1 X λ
Ln = (xm wn + bn − ymn )2 + ∥wn ∥22 (4.2)
2 2
m:rmn =1

Here, the second term is the regularization term, with λ being a positive
hyperparameter. Note that regularization is usually not applied to bn . In

15
 CHAPTER 4. OVERVIEW

practice, the average error is often used, and the loss Ln can be rewritten as:

1 X λ
Ln = (xm wn + bn − ymn )2 + ∥wn ∥22 (4.3)
2sn 2sn
m:rmn =1

Where sn is the number of items that user n has rated. In other words:

M
X
sn = rmn , (4.4)
m=1

Which is the total of the elements in column n of the rated or not matrix R.

Figure 5: Model in Python

2. Neighborhood - Collaborative Filtering
Problem: Content-based approaches ignore the preferences and behaviors
of other users. This means they miss out on potentially valuable insights
that collaborative filtering systems leverage. For instance, a collaborative
filtering system might recommend a popular book liked by people with similar
preferences, but a content-based system won’t. There are 2 branch User -
User collaborative Filtering and Item - Item collaborative Filtering
User - User collaborative Filtering (UUCF): The input of the problem is the
Utility Matrix and the most important job in UUCF must be determined to be
the same between two Users and the way to determine similarity is based on
the interest level of the items between Users. For example:

16
CHAPTER 4. OVERVIEW

Figure 6: User similarly matrix

So to conveniently calculate similarity between Users, we have to fill in the
values in the ”?” boxes. So what value should each ”?” box be filled in to
be appropriate? Many people may think of 0, but I think this will have a big
impact on making suggestions because 0 means this is a rather negative level
for the item. A value that I find more promising in this case is the average
value of the ratings that the User has rated for the remaining items. For
example, for the first user, we can calculate this average value to be 3.25,
calculated from the average of 5-3-2-2. We will calculate the utility matrix
Item - Item collaborative (IICF): Key characteristics of the user-user CF
model:
+ Relies on user similarity, using a ”Similarity matrix” to capture common
user preferences. Enables relevant recommendations for less active users.
+ Utility Y method is sparse, requiring limited user rating data to produce
relevant recommendations, based on similar users rating items similarly.
+ Similarity model can complement Utility matrix by recommending items
based on proximity to user preferences.
Additional points
+ Similarity matrix captures user similarities, enabling recommendations
for less active users.
+ With sparse Utility matrix, Similarity matrix can improve recommendations
by leveraging similar user preferences.
Item-item Collaborative Filtering is presented as a complementary approach.
For example

17
 CHAPTER 4. OVERVIEW

Figure 7: Item similarly matrix

3. Matrix Factorization
Problem: Recall that in Content-based Recommendation Systems, each
item is described by a vector x, called the item profile. In this approach, we
aim to find a vector of weights w corresponding to each user such that the
predicted rating y for an item by that user is approximated as:

y ≈ x⊤ w

Using this formulation, the Utility Matrix Y, assuming all values are filled,
is approximated as:

   
x⊤
1 w1 x⊤ ⊤
1 w 2 · · · x1 w N x⊤
1
 ⊤
x⊤ ⊤
  ⊤h
2 w2 · · · x 2 wN
 x2 w1 x  i
 =  .2  w1 w2 · · · wN = XW

Y≈
 .. .. ... ..   . 
 . . .   . 
x⊤ ⊤ ⊤
M w1 xM w2 · · · xM wN x⊤
M

where M and N denote the number of items and users, respectively.

Suppose we no longer need to predefine the item profiles x, and instead, the
feature vectors for each item can be learned simultaneously with the model
for each user (here represented by a weight vector). This means that the
optimization variables are both X and W, where:
+ X is the matrix representing all item profiles, with each row corresponding
to one item.
+ W is the matrix representing all user models, with each column corresponding
to one user.
Using this formulation, we aim to approximate the Utility Matrix Y ∈ RM ×N

18
CHAPTER 4. OVERVIEW

as the product of two matrices:

Y ≈ XW, with X ∈ RM ×K , W ∈ RK×N .

Typically, K is chosen to be much smaller than both M and N . In this case,

both matrices X and W have ranks not exceeding K . Therefore, this method
is also referred to as Low-Rank Matrix Factorization.
When X is fixed, optimizing W becomes the optimization problem in Content-
based Recommendation Systems:

N
1 X X
2 λ
L(W) = ymn − x⊤
m wn + ∥W∥2F
2s 2
n=1 m:rmn =1

When W is fixed, optimizing X reduces to the optimization problem:

N
1 X X
2 λ
L(X) = ymn − x⊤
m wn + ∥X∥2F
2s 2
n=1 m:rmn =1

19
 CHAPTER 5. NUMERICAL RESULTS

5.1 Evaluation Parameters

RMSE (Root Mean Squared Error):
v
u N
u1 X
RMSE = t (yi − ŷi )2
N
i=1

MAE (Mean Absolute Error):

N
1 X
MAE = |yi − ŷi |
N
i=1

TF-IDF
1. (Term Frequency - Inverse Document Frequency) measures the importance
of a term t in a document d within a corpus. It is calculated as:

TF-IDF(t, d) = TF(t, d) × IDF(t)

2. Term Frequency (TF)

Frequency of t in d
TF(t, d) =
Total terms in d

3. Inverse Document Frequency (IDF)

N +1
IDF(t) = log +1
1 + DF(t)

Where:
+ N : Total documents in the corpus.
+ DF(t) : Number of documents containing t

20
CHAPTER 5. NUMERICAL RESULTS

5.2 Result of Evaluating

1. Content - Based Recommendations: The Content-Based Recommendation
Model shows signs of overfitting, as evidenced by the performance gap between
training and test datasets. This suggests the model performs well on familiar
data but struggles to generalize to unseen examples. Additionally, the higher
RMSE compared to MAE on the test set indicates occasional large prediction
errors, reducing its reliability. Overall, while the model may effectively capture
individual item-user relationships, its generalization ability and robustness
remain limitations, likely requiring enhanced regularization, better feature
engineering, or a shift to more versatile recommendation approaches such as
hybrid systems.

Figure 8: Evaluation Metrics for Content-Based Recommendation

2. Neighborhood - Collaborative Filtering
User - User collaborative filtering (UUCF) is better prediction results
than content-based filtering , indicating CF effectively utilizes user relationships
to make accurate recommendations. While content-based relies solely on
product features, UUCF has the ability to explore preferences of similar users,
creating more novel recommendations. However, UUCF faces issues such as
the cold start problem and computational complexity when the number of
users is large.

21
 CHAPTER 5. NUMERICAL RESULTS

Item - Item Collaborative Filtering (IICF) The evaluation compares two

collaborative filtering approaches: UUCF and IICF. The results indicate that
IICF performs slightly better, with lower RMSE and MAE . This suggests that
IICF is more effective in reducing both large and overall prediction errors in
this dataset.

Figure 9: Evaluation Metrics for IICF

3. Matrix Factorization
GD - Gradient’s Descent
The evaluation focuses on the effectiveness of Matrix Factorization (MF)
using Gradient Descent on the MovieLens 100K dataset. Over 100 iterations,
the algorithm shows steady convergence, with the loss reducing from 5.63 to
0.24 and the RMSE on the training set decreasing from 1.18 to 0.996. The
RMSE on the test set for the Item-based MF approach is 1.0498, indicating
good predictive performance but slight overfitting. Further optimization with
hyperparameters or regularization could improve results. The predicted ratings
span a reasonable range, from around 3.09 to 4.14, which aligns well with
the actual 5-star rating scale. When comparing the predicted ratings to the
”Ratings in test set”, we can see the model is making reasonably accurate
estimates. The Prediction Ratings appear to be providing useful estimates of
the true ratings, which is the core purpose of this matrix factorization model

22
CHAPTER 5. NUMERICAL RESULTS

Figure 10: Evaluation Metrics for GD

SVD - Singular Value Decomposition
The SVD model shows RMSE values ranging from 1.01 to 1.03 as the
number of latent factors is increased from 5 to 500. Comparing this to the
Gradient Descent model you previously shared, the RMSE values were generally
higher, ranging from around 1.18 to 0.99 as the number of iterations increased.
The best RMSE from the Gradient Descent modelat 100 iterations, which is
slightly better than the lowest RMSE from the SVD model. So in summary,
the SVD model appears to be performing reasonably well, with RMSE values
in a similar range to the Gradient Descent approach. The optimal number of
latent factors for SVD was found to be 50 for this dataset.

Figure 11: Evaluation Metrics for SVD

23
 CHAPTER 6. CONCLUSIONS

6.1 Summary
In this paper, a Movie Recommendation System is developed using the MovieLens
100k dataset, which contains 100,000 movie ratings from 943 users on 1,682
movies. The system implements three key recommendation algorithms:
1. Content-based Recommendations: This algorithm recommends movies to
users based on the features of the movies and the user’s past preferences.
Features such as genre, director, and cast are used to calculate similarities
between movies.
2. Neighborhood-based Collaborative Filtering: This algorithm makes predictions
based on the preferences of similar users. It identifies users who have similar
movie preferences and recommends movies that those users have rated highly.
3. Matrix Factorization (e.g., Singular Value Decomposition - SVD): Matrix
factorization techniques break down the user-item interaction matrix into lower-
dimensional matrices representing latent features of users and movies. This
method captures hidden patterns in user preferences and movie characteristics
that are not immediately apparent from the data.
Dataset and Algorithm Challenges
While the MovieLens 100k dataset is a valuable resource, it has some inherent
limitations. Sparsity: The dataset is sparse, meaning many users have rated only
a small fraction of available movies, which makes it difficult to make accurate
predictions for unseen items. Cold Start Problem: New users or new movies
without sufficient ratings make it challenging to generate meaningful recommendations.
This issue affects all three algorithms. Bias in Data: The dataset contains bias in
terms of user activity (some users rate far more movies than others), which can
impact the fairness of recommendations.
6.2 Suggestion for Future Works
To enhance the Movie Recommendation System, integrating Deep Learning
models can help to better capture complex patterns in user preferences and item
relationships. Here are several deep learning approaches that can improve the
system:
1. Deep Neural Networks (DNNs): DNNs can learn latent representations of
users and items by encoding user-item interactions. This can overcome some
limitations of traditional methods like matrix factorization, particularly in

24
CHAPTER 6. CONCLUSIONS

capturing non-linear relationships between users and movies.

2. Autoencoders: Autoencoders can be used to reconstruct the ratings matrix by
learning compressed representations of user-item interactions. This approach
helps deal with data sparsity and improves the system’s ability to generalize
over unseen ratings.
3. Recurrent Neural Networks (RNNs): For sequential recommendations, RNNs,
particularly LSTMs and GRUs, can capture the temporal dynamics of a user’s
preferences, learning from the order of previously watched movies to predict
what the user might want to watch next.
4. Deep Reinforcement Learning (DRL): DRL treats the recommendation task
as an optimization problem, where the system learns to make decisions based
on user interactions and feedback. This approach is useful for continuously
improving recommendations based on real-time user engagement.
5. Hybrid Models with Neural Collaborative Filtering (NCF): By combining
traditional collaborative filtering techniques with deep learning models, such
as DNNs, hybrid models can capture both linear and non-linear user-item
interactions, offering more personalized and effective recommendations.
Benefits of Deep Learning Models
1. Improved Accuracy: Deep learning can capture more complex patterns, leading
to more precise recommendations.
2. Better Adaptability: Models like RNNs and DRL can dynamically adapt to
changing user preferences over time.
3. Scalability: Deep learning models can scale to large datasets, making them
suitable for systems with millions of users and items.
Challenges
1. Computational Complexity: Training deep learning models can be resource-
intensive, requiring significant computational power.
2. Data Requirements: These models need large datasets to train effectively
and avoid overfitting.
3. Interpretability: Deep learning models, particularly DNNs and GNNs, are
often seen as “black boxes” and are harder to interpret compared to traditional
recommendation methods.

25
 REFERENCE

Vu Huu Tiep provides the ”Machine Learning Co Ban” website accessed at

https://machinelearningcoban.com/.
The GeeksforGeeks article on ”What are Recommender Systems?” accessed at
https://www.geeksforgeeks.org/what-are-recommender-sys
tems/.
The Wikipedia article on ”Recommender Systems” at https://en.wikip
edia.org/wiki/Recommender_system
Roughgarden, T., & Valiant, G. (2024). CS168: The Modern Algorithmic
Toolbox Lecture #9: The Singular Value Decomposition (SVD) and Low-Rank
Matrix Approximations., accessed at https://web.stanford.edu/class
/cs168/l/l9.pdf/.
Roughgarden, T., & Valiant, G. (2024). CS168: The Modern Algorithmic
Toolbox Lecture #9: The Singular Value Decomposition (SVD) and Low-Rank
Matrix Approximations., accessed at https://web.stanford.edu/class
/cs168/l/l9.pdf.
Rahul Makhijani, Saleh Samaneh, Megh Mehta. CS229 Lecture Discussion:
Collaborative Filtering Recommender Systems, accessed at https://cs229.
stanford.edu/proj2014/Rahul%20Makhijani,%20Saleh%20Sam
aneh,%20Megh%20Mehta,%20Collaborative%20Filtering%20Re
commender%20Systems.pdf
Zheng Wen. CS229 (2008): Recommendation System Based on Collaborative
Filtering, accessed at https://cs229.stanford.edu/proj2008/We
n-RecommendationSystemBasedOnCollaborativeFiltering.p
df

26