Copyright (c) 2019 Sebastian Raschka

https://github.com/rasbt/python-machine-learning-book-3rd-edition

Chapter 1 - Giving Computers the Ability to Learn from Data

Overview

• Building intelligent machines to transform data into knowledge
• The three different types of machine learning
  • Making predictions about the future with supervised learning
    • Classification for predicting class labels
    • Regression for predicting continuous outcomes
  • Solving interactive problems with reinforcement learning
  • Discovering hidden structures with unsupervised learning
    • Finding subgroups with clustering
    • Dimensionality reduction for data compression
    • An introduction to the basic terminology and notations
• A roadmap for building machine learning systems
  • Preprocessing - getting data into shape
  • Training and selecting a predictive model
  • Evaluating models and predicting unseen data instances
• Using Python for machine learning
• Installing Python packages
• Summary

Machine Learning (ML)

• a subfield of Artificial Integlligence (AI)
• application and science of algorithms that make sense of data
• one of the most exciting fields in the CS!
• with abundant of data (~ 2MB data created per second per person!), ML algorithms can be used to spot patterns in data and make predictios about the future events

Building intelligent machines/algorithms to transform data into knowledge

• data is abundant in both structured and unstructed form
• ML involves self-learning algorithms that derive knowledge from data in order to make predictions
• ML applications are already ubiquitous:
  • spam filter
  • web search engines
  • Network intrusion detection and prevention system
  • digital assistants (Apple Siri, Amazon Alexa, Google Assistant, Microsoft Cortana, etc.)
  • self-driving cars (Google, Uber, Tesla, etc.)
  • skin cancer detection (https://www.nature.com/articles/nature21056)
  • AlphaZero from DeepMind has beaten human champions in Chess, Go and Shogi (Japanese Chess)

Three different types of machine learning

Supervised learning

• goal is to learn a model from labeled training data that allows us to make predictions about unseen/unlabeled data

• two types: classification and regression

Classification

• classify samples/data to fixed discreted class (labels)
• can be binary class classification
  • e.g. labeling emails as spam or ham, classify internet traffic as malicious or benign, classify programs as malware or benign, classify tumor or benign, etc.
• can bee multi-class classification
  • classify type of malware (adware, spyware, trojans, keyloggers, rootkits, etc.)
  • classify type various stages of cancer (stage 0-4)
  • classify images of people, etc.
• the following figure depicts binary classification problem between A and B samples
  

Regression

• the outcome is continous value
  • e.g. predicting stock prices, home prices, weather prediction (max/min temps, wind speed, humidity), etc.
• the following figure depicts predicting continuous outcomes given some data
• given a feature variable, $x$, and a targe variable $y$, we fit a straight line to this data that minimizes the distance between the data points and the fitted line

Reinforcement Learning

• goal is to develop a system (agent) that improves its performance based on interactions with the environment

  • loosely speaking, this is AI
  • similar to how humans and animals with intelligence learn

• the heart of the system is so-called reward signal

  • reward can be positive (good) or negative (bad)
  • e.g. training dogs by giving rewards

• the agent interacts with the environment and learn a series of actions by maximizing the rewards via trial-and-error or planning

• the following figure depicts reinforcement learning

Unsupervised Learning

• deal with unlabelled data or data of unknown structure
• technique can be used to extract meaningful information without the guidance of a known outcome variable or reward function

Finding subgroups with clustering

• finding natural or meaningful subgroups (clusters) without having any prior knowledge
• the following figure illustrates clustering of data into 3 sub-groups

Basic terminology and notations

Dataset and representation

• most ML algorithms learn from data
• classic example of machine learning dataset is the Iris dataset
• contains the measurements of 150 Iris flowers from three species: Setosa, Versicolor and Virginica
• the measuresments are also called the features
• dataset is typically 2-dimensional matrix
• each row represents a sample/observation/instance and each column is a feature value for that sample
• the following figure represents a slice of the Iris dataset

Irisi Dataset

• Iris dataset with 150 samples and four features can be written as a $150x4$ matrix:

  $ \begin{equation*} \textbf{X} \in {\mathbb{R}}^{150x4} : \textbf{X}^{150x4} = \begin{bmatrix} x_{1}^{1} & x_2^{1} & x_3^{1} & x_4^{1} \ x_{1}^{2} & x_2^{2} & x_3^{2} & x_4^{2} \ \vdots & \vdots & \vdots & \vdots \ x_{1}^{150} & x_2^{150} & x_3^{150} & x_4^{150} \end{bmatrix} \end{equation*} $

Notational Conventions

• List of Mathematical Symbols: https://oeis.org/wiki/List_of_LaTeX_mathematical_symbols
• superscript $i$ refers to the $i^{th}$ sample/example
• subscript $j$ refers to the $j^{th}$ dimension/feature of the sample
• lowercase, bold-face letters refer to a single vector $\textbf{x} \in \mathbb{R}^{1xm}$
• uppercase, bold-face letters refer to a matrix $\textbf{X} \in \mathbb{R}^{nxm}$
• single element in a vector represented in italics: $\textit{x}_{m}$
• single element in a matrix represented in italics: $\textit{x}_m^{(n)}$

E.g.

• $x_1^{(150)}$ refers to the first dimension of flower sample #150 (sepal length)

• $i^{th}$ flower sample in the matrix can be written as a 4-dimensional row vector:

  $ \textbf{x}^{(i)} \in \mathbb{R}^{1x4}: \textbf{x}^{(i)} = \begin{bmatrix} x_1^{(i)} & x_2^{(i)} & x_3^{(i)} & x_4^{(i)}\end{bmatrix}$

• $j^{th}$ feature in matrix is a 150-dimensional column vector:

  $\begin{equation*} \textbf{x}{j} \in \mathbb{R}^{150x1}: \textbf{x}{j} = \begin{bmatrix} x_j^{(1)} \ x_j^{(2)} \ \vdots\ x_j^{(150)} \end{bmatrix} \end{equation*} $

• target variables is 150-dimensional column vector:

$\textbf{y} = \begin{bmatrix} y^{(1)} \ y^{(2)} \ \vdots \ y^{(150)} \end{bmatrix} (y \in {Setosa, Versicolor, Virginica}) $

Training sample

• a row in a table representating the dataset also called observation, record, instance

Training

• Model fitting, similar to parameter estimation

Feature

• a column in a data table (matrix), also called variable, attribute or covariate, input

Target

• also called outcome, class, label, response variable, ground truth, dependent variable

Loss function

• also called cost function, error function

Public Repository of Machine Learning Datasets

1. UCI Machine Learning Repository - https://archive.ics.uci.edu/ml/datasets.php
2. UNB - Canadian Institute for Cybersecurity Datasets - https://www.unb.ca/cic/datasets/index.html
3. Kaggle Public Datasets - https://www.kaggle.com/datasets
4. Sci-kit learn Real-World Datasets - https://scikit-learn.org/stable/datasets/real_world.html
5. Seaborn Datasets - https://github.com/mwaskom/seaborn-data
6. https://www.openml.org
7. https://huggingface.co/

A roadmap for building machine learning systems

• the following figure shows a typical workflow for using machine learning in predictive modeling

Preprocessing - getting data into shape

• raw data can be structured (csv, xml, json, database, etc.) or unstructured (text, images, videos, audios, etc.)
• it's important to convert raw data into feature vector by extracting appropriate features
  • e.g. in Iris dataset, sepal and petal length and width
• feature values are preferred to be on the same scale for optimal performance typically in the range [0, 1]
  • standard normal distribution with zero mean and unit variance (covered in later chapters)
• features may be highly correlated and therefore redundant to a certain degree
  • feature selection and dimensionality reduction techniques are used
  • requires less memory and CPU time
  • may improve predictive performance of model
  • reduce noise (remove irrelevant features)
• dataset is typically divided into two parts (training set and testing/validation set)
• training set is used to train the model
• testing/validation set is used to evaulate the model to see how it would perform or generalize to new data
• stratified cross-validation techniques are quite common as well when dataset is scarce (small)
  • e.g. 10-fold cross-validation is a common practice

Training and selecting a predictive model

• each classifier or ML algorithm may have its inherent biases on different type of problem and dataset
• in practice, it's essential to comapre at least a handful of different algorithms in order to train and select the best performing model
• typically certain pre-defined metrics are used to compare performance
  • e.g. accuracy (proportion of correctly classified instances) and many others (covered later...)
• default parameters used by algorithms may work well but not guaranteed
• these hyperparameters may need to be fine-tuned for optimal performance

Evaluating models and predicting unseen data instances

• the best model, that has been fitted on the training set, is evaluated against the unseen data to estimate generalization error
• if satisfied with the model's performance, it's then deployed and applied to the new, unknown dataset in the real-word
• same scaling techniques (normalization) and dimensionality reduction technique and the parameters are latter reapplied to transform the new data instances

Why Python for machine learning

• active developers and open source community producing a large number of useful libraries
• NumPy, SciPy on lower-layer uses Fortran and C implementations for fast vector operations on multidimensional arrays
• scikit-learn library has all the popular open-soruce machine learning algorithms
• several deep learning frameworks (TensorFlow, Fast.ai, PyTorch) have Python extensions or built with Python