Introduction to Machine Learning

Md. Abid Hasan Miazy Md. Montasir Mamun sagor

ID: 12008001 ID: 12008025
Deptartment of Computer Science and Engineering Department of Computer Sceince and Engineering
Comilla University Comilla University
Cumilla, Bangladesh Cumilla, Bangladesh
abidsusma4288miazy@gmail montasirsagor0@gmail.com

Md. Mehedi Hasan Parvez Romjan Chowodhury

ID: 11908015 ID: 11908048
Department of Computer Science and Engineering Department of Computer Science and Engineering
Comilla Universilty Comilla University
Cumilla, Bangladesh Cumilla, Bangladesh
mehediparvez11908015@stud.cou.ac.bd mrchowdhury587@gmail.com

Abstract—Machine learning is one of the subfields of artificial Key Characteristics:

intelligence which focuses on the development of systems that
are capable of automatically learning from data and making • Labeled Data: Every training data has a corresponding
predictions or decisions. In this work, we explain the main types label.
of machine learning - supervised, unsupervised and reinforcement • Objective: Minimize the difference between expected
learning – discussing their algorithms and applications in detail. output and actual output.
Moreover, we cover essential topics such as underfitting and • Applications: Identifying spam, classifying images and
overfitting, empirical vs. expected risk, and we also describe the
machine learning pipeline. This type of knowledge provides the giving medical diagnosis.
groundwork to understand how machine learning techniques can Example: For the purpose of supervised learning, let us
and should be optimized for practical applications. consider an example where emails are to be classified as
Index Terms—Machine Learning, Supervised Learning, Unsu-
pervised Learning, Reinforcement Learning, Overfitting, Under- ’spam’ or ’not spam’. The model is executed using a dataset
fitting, Empirical Risk, Expected Risk, ML Pipeline. containing all emails labeled ‘spam’ or ‘not spam’. After
training, the model understands the features of emails (like
keywords or sender information) and is able to accurately
I. I NTRODUCTION
classify future emails.
Identifying patterns, and designing algorithms to process Common Algorithms:
increasingly large volumes of data are some of the defining • Linear Regression
features of the modern world. In fact, machine learning is • Logistic Regression
a branch of artificial intelligence which allows computers to • Support Vector Machines (SVM)
learn from information without explicit human interaction, • Decision Trees
hence making it increasingly crucial in contemporary society. • Random Forests
This report presents the basic concepts of ML along with its • k-Nearest Neighbors (k-NN)
differentiated types focusing on supervised, unsupervised, and • Naı̈ve Bayes
reinforcement learning. Moreover, overfitting and underfitting
along with examples will be covered as well as definitions of Applications:
empirical and expected risks will be provided. • Healthcare: Predicting disease progression given patient
health records.
II. S UPERVISED , U NSUPERVISED , AND R EINFORCEMENT • Finanacial: Estimating the credit score given a customer’s
L EARNING financial history.
• Retail: Forecasting demand in order to estimate the stock
A. Supervised Learning needed.
Supervised learning is similar to learning with the aid of a
teacher. In this case, the model is taught using a labeled dataset
B. Unsupervised Learning
where every piece of data is matched with the relevant output.
The main focus is to help the model understand how to map Data which has no labeled information can be grouped
inputs and outputs so that it can one day generate predictions or strucutre in unsupervised machine learning methods and
for novel data. models.
 Fig. 2. Workflow of Unsupervised Machine Learning

Fig. 1. Workflow of Supervised Machine Learning

Key Characteristics:
• Raw Data: No prior defined labels or marks.
• Objective: unmask underlying structures and arrang-
ments.
Example: Grouping customer behavior. A machine can
group people based on their spending habits without knowing
the real grouping..
Common Algorithms:
• K mean cluster or simply K means
Fig. 3. Workflow of Reinforcement Machine Learning
• Hierarchial cluster
• Density Based Spatial Clustering of Applications with
Noise (DBSCAN) III. U NDERFITTING AND OVERFITTING
• PCA Principal componenet or PCA A. Underfitting
• T-distributed Stochastic Neighbor Embedding or t-SNE
When the model does not incorporate the underlying data
• Auto-encoders
complexities and is overly simplistic, it leads to underfitting.
Applications: Both the training data and test data yield disappointing results.
• Marketing: make mesured ad campaigns Characteristics:
• Healthcare: Mark abnormalities in imaging data
• High Bias: The model lacks nuance, having a consistent
• Economic: Help determine false and true claims
unrealistic view of the data
C. Reinforcement Learning • Low Variance: The structure which guides the model is

Reinforcement learning in training an agent to make choices uniformly inaccurate

• The model underperforms for both training and test
decisions by interacting in an environment and learning
through rewards and penalties. datasets.
Key Characteristics: Example: Trying to estimate a property’s value using only
• Interaction: It includes an agent that acts and receives the square footage and zeroing in on its location or number
feedback. of rooms is an example of underfitting.
• Objective: Maximal cumulative reward. Prevention Techniques:
• Applications: Robotics, gaming, self-driving cars. • Withstand the irritation of patterns that are clearly non-

Example: A robot navigating a maze receives rewards for linear by adding layers to the model.
reaching the goal and penalties for hitting walls learning the • Include more relevant features to provide the model with

optimal path over time.. additional information.

Common Algorithms: B. Overfitting
• Q-Learning
Overfitting is a phenomenon in machine learning where a
• Deep Q-Networks (DQN)
model learns patterns too well, including random noise, which
• Policy Gradient Methods
leads to excellent performance within the training set but poor
• Actor-Critic Methods
outcomes on new test datasets because the model is unable to
• Proximal Policy Optimization (PPO)
make generalizations.
Applications: Characteristics:
• Gaming: Chess, Go
• Minimal Bias: The model does not show significant
• Robotics: Object manipulation
deviations from the actual values due to details captured
• Finance: Adaptive trading strategies
in the training data.
 • High Variance: The model produces different outputs for A. Pipeline Stages
the same function if applied to different datasets. To make predictions, data passes through several steps to
• Great Results During Training: The model has undergone
produce a machine learning model capable of making accurate
extensive training and shows trivial error on the training predictions. These steps include:
set.
• Poor Results on Testing: The model has trained too
1) Data Collection: Data is collected from different
exclusively tailored to the training data, therefore will sources such as databases, APIs, or files. This data is
underperform when tested against any unfamiliar data often raw and requires preprocessing.
2) Data Pre-processing: Involves handling missing values,
Example: A highly sophisticated decision tree that memo- removing duplicates, and correcting errors to improve
rizes every possible pattern for each training instance might the model’s predictions. This step often requires domain
achieve perfect accuracy during training. Such a tree is likely knowledge and creativity.
to underperform on real-world scenarios filled with slight 3) Feature Engineering: Creating new features or select-
deviations when presented with previously unseen data. ing relevant features to enhance model performance.
Prevention Techniques: 4) Data Splitting: Dividing the data into training and
• Cross-validation testing sets to evaluate model performance.
• Regularization 5) Model Selection: Choosing an appropriate machine
• Pruning learning algorithm based on the problem type (e.g.,
• Ensemble methods classification, regression), data characteristics, and per-
formance needs.
IV. E MPIRICAL R ISK AND E XPECTED R ISK
6) Model Training: Training the model on the training
A. Empirical Risk dataset using the selected algorithm.
Empirical risk refers to the average loss of a model over a 7) Model Evaluation: Assessing the model’s performance
finite sample of training data. Given a dataset {(xi , yi )}ni=1 using a separate testing dataset or through cross-
drawn i.i.d. from the underlying data distribution D, the validation techniques.
empirical risk R̂n (h) is defined as: 8) Model Tuning: Adjusting hyperparameters to improve
the model’s performance.
n
1X 9) Model Deployment: Integrating the trained model into a
R̂n (h) = L(h(xi ), yi )
n i=1 production environment where it can make predictions
on unseen data. This may involve creating APIs and
Empirical risk serves as an approximation of the true integrating with other systems.
(expected) risk based on observed data. The principle of 10) Monitoring and Maintenance: Continuously monitor-
Empirical Risk Minimization (ERM) in machine learning ing the model in production and updating it as needed
involves selecting a hypothesis h ∈ H that minimizes the to maintain performance.
empirical risk.
B. Benefits of Machine Learning Pipelines
B. Expected Risk
• Modularization: Simplifies development by breaking the
The expected risk, also known as the true or real risk, is process into independent steps.
a theoretical measure of how well a model performs over the • Efficiency: Automates repetitive tasks, saving time and
entire data distribution, including unseen cases. It represents effort.
the anticipated value of the loss function based on the joint • Scalability: Handles large datasets effectively.
distribution of input-output pairs (x, y). • Experimentation: Facilitates experimentation with dif-
For a hypothesis (model) h and a loss function L(h(x), y), ferent techniques and configurations.
the expected risk R(h) is defined as: • Deployment: Streamlines transition from development to
Z production.
R(h) = E(x,y)∼D [L(h(x), y)] = L(h(x), y) dD(x, y) • Collaboration: Enhances teamwork by structuring the
workflow.
Here, D denotes the (unknown) true probability distribution • Version Control and Documentation: Enables tracking
of the data. Since D is usually not known in practice, the of changes in code and configurations.
expected risk cannot be computed exactly.
VI. C ONCLUSION
V. M ACHINE L EARNING P IPELINE Understanding the core differences among the machine
A machine learning pipeline is a series of interconnected learning algorithms is crucial for selecting the proper way to
data processing and modeling steps designed to automate, solve the problem. Also properly recognizing the overfitting
standardize, and streamline the process of building, training, and underfitting and resolving the issues gives insights to build
evaluating, and deploying machine learning models. suitable models.
 ACKNOWLEDGMENT
The authors would like to express their sincere gratitude
to Dr. Mahmudul Hasan, Head of the Department of Com
puter Science and Engineering at Comilla University, for
his invaluable support in the conception and design of this
study. His insightful guidance and encouragement significantly
contributed to the development and direction of this research.
R EFERENCES
[1] S. Badillo, B. Banfai, F. Birzele, I. I. Davydov, et al., “An introduction
to machine learning,” *Clinical Pharmacology Therapeutics*, Wiley
Online Library, 2020.
[2] A. Smola and S. V. N. Vishwanathan, “Introduction to machine
learning,” *Cambridge University*, UK, 2008. [Online]. Available:
academia.edu
[3] Y. Baştanlar and M. Özuysal, “Introduction to machine learning,” in
*miRNomics: MicroRNA biology and applications*, Springer, 2013.
[4] M. Kubat, *An Introduction to Machine Learning*, Springer, 2017.
[5] N. J. Nilson, “Introduction to machine learning,” Draft Textbook, 1998.
[Online]. Available: bafflerbach.github.io
[6] G. Rebala, A. Ravi, and S. Churiwala, *An Introduction to Machine
Learning*, Springer, 2019.
[7] E. Alpaydin, *Introduction to Machine Learning*, MIT Press, 2020.
[8] R. Y. Choi, A. S. Coyner, et al., “Introduction to machine learning, neural
networks, and deep learning,” *Investigative Ophthalmology Visual
Science*, 2020.
[9] O. Simeone, “A brief introduction to machine learning for engineers,”
*Foundations and Trends in Signal Processing*, vol. 12, no. 3–4, pp.
200–399, 2018.
[10] S. Lemm, B. Blankertz, T. Dickhaus, and K. R. Müller, “Introduction
to machine learning for brain imaging,” *NeuroImage*, vol. 56, no. 2,
pp. 387–399, 2011.