<!-- Slide number: 19 -->

Unsupervised Learning

Data Feature

Feature 1

...

Feature n

Internal similarity

Unsupervised learning algorithm

...

Feature n

Feature 1

Feature 1

Feature n

...

| Monthly Consumption | Commodity | Consumption Time |

| --- | --- | --- |

| 1000–2000 | Badminton racket | 6:00–12:00 |

| 500–1000 | Basketball | 18:00–24:00 |

| 1000–2000 | Game console | 00:00–6:00 |

| Category |

| --- |

| Cluster 1 |

| Cluster 2 |

Notes:

<!-- Slide number: 20 -->

Unsupervised Learning - Clustering Questions

Clustering: classifies samples in a sample dataset into several categories based on the clustering model. The similarity of samples belonging to the same category is high.

Which audiences like to watch movies of the same subject?

Which of these components are damaged in a similar way?

Notes:

<!-- Slide number: 21 -->

Semi-Supervised Learning

Data Feature

Label

Feature 1

Goal

...

Feature n

Semi-supervised learning algorithms

...

Unknown

Feature n

Feature 1

Feature 1

Feature n

...

Unknown

| Weather | Temperature | Wind Speed |

| --- | --- | --- |

| Sunny | Warm | Strong |

| Rainy | Cold | Fair |

| Sunny | Cold | Weak |

| Enjoy Sports |

| --- |

| Yes |

| / |

| / |

Notes:

<!-- Slide number: 22 -->

Reinforcement Learning

The model perceives the environment, takes actions, and makes adjustments and choices based on the status and award or punishment.

Model

Environment

Notes:

Reinforcement learning uses a series of actions to maximize the reward function to learn models.

Both good and bad behaviors can help reinforcement learning in model learning.

For example, autonomous vehicles learn by continuously interacting with the environment.

<!-- Slide number: 23 -->

Reinforcement Learning - Best Behavior

Reinforcement learning: always looks for best behaviors. Reinforcement learning is targeted at machines or robots.

Autopilot: Should it brake or accelerate when the yellow light starts to flash?

Cleaning robot: Should it keep working or go back for charging?

Notes:

<!-- Slide number: 24 -->

Machine learning algorithm

Machine Learning Classification

Machine Learning Process

Other Key Machine Learning Methods

Common Machine Learning Algorithms

Case study

Notes:

<!-- Slide number: 25 -->

Machine Learning Process

Model deployment and integration

Feature extraction and selection

Model evaluation

Data collection

Data cleansing

Model training

Feedback and iteration

Notes:

<!-- Slide number: 26 -->

Basic Machine Learning Concept — Dataset

Dataset: a collection of data used in machine learning tasks. Each data record is called a sample. Events or attributes that reflect the performance or nature of a sample in a particular aspect are called features.

Training set: a dataset used in the training process, where each sample is referred to as a training sample. The process of creating a model from data is called learning (training).

Test set: Testing refers to the process of using the model obtained after learning for prediction. The dataset used is called a test set, and each sample is called a test sample.

Notes:

<!-- Slide number: 27 -->

Checking Data Overview

Typical dataset form

Feature 3

Feature 1

Feature 2

Label

| No. | Area | School Districts | Direction | House Price |

| --- | --- | --- | --- | --- |

| 1 | 100 | 8 | South | 1000 |

| 2 | 120 | 9 | Southwest | 1300 |

| 3 | 60 | 6 | North | 700 |

| 4 | 80 | 9 | Southeast | 1100 |

| 5 | 95 | 3 | South | 850 |

Training set

Test set

Notes:

<!-- Slide number: 28 -->

Importance of Data Processing

Data is crucial to models. It is the ceiling of model capabilities. Without good data, there is no good model.

Data preprocessing

Data cleansing

Data normalization

Normalize data to reduce noise and improve model accuracy.

Fill in missing values, and detect and eliminate causes of dataset exceptions.

Data dimension reduction

Simplify data attributes to avoid dimension explosion.

Notes:

<!-- Slide number: 29 -->

Workload of Data Cleansing

Statistics on data scientists' work in machine learning

3% Remodeling training datasets

5% Others

4% Optimizing models

9% Mining modes from data

19% Collecting datasets

60% Cleansing and sorting data

CrowdFlower Data Science Report 2016

Notes:

<!-- Slide number: 30 -->

Data Cleansing

Most machine learning models process features, which are usually numeric representations of input variables that can be used in the model.

In most cases, the collected data can be used by algorithms only after being preprocessed. The preprocessing operations include the following:

Data filtering

Processing of lost data

Processing of possible exceptions, errors, or abnormal values

Combination of data from multiple data sources

Data consolidation

Notes:

<!-- Slide number: 31 -->

Dirty Data (1)

Generally, real data may have some quality problems.

Incompleteness: contains missing values or the data that lacks attributes

Noise: contains incorrect records or exceptions.

Inconsistency: contains inconsistent records.

Notes:

<!-- Slide number: 32 -->

Dirty Data (2)

| # | Id | Name | Birthday | Gender | IsTeacher | #Students | Country | City |

| --- | --- | --- | --- | --- | --- | --- | --- | --- |

| 1 | 111 | John | 31/12/1990 | M | 0 | 0 | Ireland | Dublin |

| 2 | 222 | Mery | 15/10/1978 | F | 1 | 15 | Iceland | |

| 3 | 333 | Alice | 19/04/2000 | F | 0 | 0 | Spain | Madrid |

| 4 | 444 | Mark | 01/11/1997 | M | 0 | 0 | France | Paris |

| 5 | 555 | Alex | 15/03/2000 | A | 1 | 23 | Germany | Berlin |

| 6 | 555 | Peter | 1983-12-01 | M | 1 | 10 | Italy | Rome |

| 7 | 777 | Calvin | 05/05/1995 | M | 0 | 0 | Italy | Italy |

| 8 | 888 | Roxane | 03/08/1948 | F | 0 | 0 | Portugal | Lisbon |

| 9 | 999 | Anne | 05/09/1992 | F | 0 | 5 | Switzerland | Geneva |

| 10 | 101010 | Paul | 14/11/1992 | M | 1 | 26 | Ytali | Rome |

Missing value

Invalid value

Value that should be in another column

Invalid duplicate item

Misspelling

Incorrect format

Attribute dependency

Notes:

<!-- Slide number: 33 -->

Data Conversion

After being preprocessed, the data needs to be converted into a representation form suitable for the machine learning model. Common data conversion forms include the following:

With respect to classification, category data is encoded into a corresponding numerical representation.

Value data is converted to category data to reduce the value of variables (for age segmentation).

Other data

In the text, the word is converted into a word vector through word embedding (generally using the word2vec model, BERT model, etc).

Process image data (color space, grayscale, geometric change, Haar feature, and image enhancement)

Feature engineering

Normalize features to ensure the same value ranges for input variables of the same model.

Feature expansion: Combine or convert existing variables to generate new features, such as the average.

Notes:

<!-- Slide number: 34 -->

Necessity of Feature Selection

Generally, a dataset has many features, some of which may be redundant or irrelevant to the value to be predicted.

Feature selection is necessary in the following aspects:

Notes:

<!-- Slide number: 35 -->

Feature Selection Methods - Filter

Filter methods are independent of the model during feature selection.

By evaluating the correlation between each feature and the target attribute, these methods use a statistical measure to assign a value to each feature. Features are then sorted by score, which is helpful for preserving or eliminating specific features.

Common methods

Pearson correlation coefficient

Chi-square coefficient

Mutual information

Limitations

The filter method tends to select redundant variables as the relationship between features is not considered.

Select the optimal feature subset

Traverse all features

Train models

Evaluate the performance

Procedure of a filter method

Notes:

<!-- Slide number: 36 -->

Feature Selection Methods - Wrapper

Wrapper methods use a prediction model to score feature subsets.

Wrapper methods consider feature selection as a search issue for which different combinations are evaluated and compared. A predictive model is used to evaluate a combination of features and assign a score based on model accuracy.

Common methods

Recursive feature elimination (RFE)

Limitations

Wrapper methods train a new model for each subset, resulting in a huge number of computations.

A feature set with the best performance is usually provided for a specific type of model.

Select the optimal feature subset

Generate a feature subset

Traverse all features

Train models

Evaluate models

Procedure of a wrapper method

Notes:

<!-- Slide number: 37 -->

Feature Selection Methods - Embedded

Embedded methods consider feature selection as a part of model construction.

The most common type of embedded feature selection method is the regularization method.

Regularization methods are also called penalization methods that introduce additional constraints into the optimization of a predictive algorithm that bias the model toward lower complexity and reduce the number of features.

Common methods

Lasso regression

Ridge regression

Select the optimal feature subset

Generate a feature subset

Train models

+ Evaluate the effect

Traverse all features

Procedure of an embedded method

Notes:

The penalty attribute of the model is highlighted, which is used to eliminate unimportant features.

<!-- Slide number: 38 -->

Overall Procedure of Building a Model

Model Building Procedure

1 2 3

Data splitting:

Divide data into training sets, test sets, and validation sets.

Model training:

Use data that has been cleaned up and feature engineering to train a model.

Model verification:

Use validation sets to validate the model validity.

6

5

4

Model fine-tuning:

Continuously tune the model based on the actual data of a service scenario.

Model deployment:

Deploy the model in an actual production scenario.

Model test:

Use test data to evaluate the generalization capability of the model in a real environment.

Notes:

After data cleansing and feature extraction, we need to start building the model. The general procedure for building a model is shown above (supervised learning).

<!-- Slide number: 39 -->

Examples of Supervised Learning - Learning Phase

Use the classification model to predict whether a person is a basketball player.

Feature (attribute)

Target

| Name | City | Age | Label |

| --- | --- | --- | --- |

| Mike | Miami | 42 | yes |

| Jerry | New York | 32 | no |

| Bryan | Orlando | 18 | no |

| Patricia | Miami | 45 | yes |

| Elodie | Phoenix | 35 | no |

| Remy | Chicago | 72 | yes |

| John | New York | 48 | yes |

Service data

Training set

The model searches for the relationship between features and targets.

(Cleansed features and tags)

Splitting

Task: Use a classification model to predict whether a person is a basketball player under a specific feature.

Test set

Use new data to verify the model validity.

Model training

Each feature or a combination of several features can provide a basis for a model to make a judgment.

Notes:

Model example

<!-- Slide number: 40 -->

Examples of Supervised Learning - Prediction Phase

| Name | City | Age | Label |

| --- | --- | --- | --- |

| Marine | Miami | 45 | ? |

| Julien | Miami | 52 | ? |

| Fred | Orlando | 20 | ? |

| Michelle | Boston | 34 | ? |

| Nicolas | Phoenix | 90 | ? |

Unknown data

Recent data, it is not known whether the people are basketball players.

New data

IF city = Miami → Probability = +0.7

IF city= Orlando → Probability = +0.2

IF age > 42 → Probability = +0.05*age + 0.06

IF age ≤ 42 → Probability = +0.01*age + 0.02

Application model

| Name | City | Age | Prediction |

| --- | --- | --- | --- |

| Marine | Miami | 45 | 0.3 |

| Julien | Miami | 52 | 0.9 |

| Fred | Orlando | 20 | 0.6 |

| Michelle | Boston | 34 | 0.5 |

| Nicolas | Phoenix | 90 | 0.4 |

Possibility prediction

Apply the model to the new data to predict whether the customer will change the supplier.

New data

Prediction data

Notes:

Model example

The model here can be considered similar to the decision tree model.

<!-- Slide number: 41 -->

What Is a Good Model?

Generalization capability

Can it accurately predict the actual service data?

Interpretability

Is the prediction result easy to interpret?

Prediction speed

How long does it take to predict each piece of data?

Practicability

Is the prediction rate still acceptable when the service volume increases with a huge data volume?

Notes:

Which factors are used to determine a model?

The last three are engineering factors. The generalization capability is the most important factor.

<!-- Slide number: 42 -->

Model Validity (1)

Generalization capability: The goal of machine learning is that the model obtained after learning should perform well on new samples, not just on samples used for training. The capability of applying a model to new samples is called generalization or robustness.

Error: difference between the sample result predicted by the model obtained after learning and the actual sample result.

Training error: error that you get when you run the model on the training data.

Generalization error: error that you get when you run the model on new samples. Obviously, we prefer a model with a smaller generalization error.

Underfitting: occurs when the model or the algorithm does not fit the data well enough.

Overfitting: occurs when the training error of the model obtained after learning is small but the generalization error is large (poor generalization capability).

Notes:

Once the form of a problem's hypothesis is given, all possible functions constitute a space, which is hypothesis space. The problem of machine learning is searching for a suitable fitting function in a hypothesis space.

Overfitting: It occurs frequently in complex mathematical models. To prevent overfitting, we can simplify mathematical models, end training before overfitting, or use dropout/weight decay methods.

Underfitting: It occurs if the mathematical model is too simple or the training time is too short. To solve the underfitting problem, use a more complex model or extend the training time.

<!-- Slide number: 43 -->

Model Validity (2)

Model capacity: model's capability of fitting functions, which is also called model complexity.

When the capacity suits the task complexity and the amount of training data provided, the algorithm effect is usually optimal.

Models with insufficient capacity cannot solve complex tasks and underfitting may occur.

A high-capacity model can solve complex tasks, but overfitting may occur if the capacity is higher than that required by a task.


Underfitting

Not all features are learned.

Good fitting

Overfitting

Noises are learned.

Notes:

The effective capacity is restricted by algorithms, parameters, and regularization.

<!-- Slide number: 44 -->

Overfitting Cause — Error

Total error of final prediction = Bias2 + Variance + Irreducible error

Generally, the prediction error can be divided into two types:

Error caused by "bias"

Error caused by "variance"

Variance:

Offset of the prediction result from the average value

Error caused by the model's sensitivity to small fluctuations in the training set

Bias:

Difference between the expected (or average) prediction value and the correct value we are trying to predict.

Variance

Bias

Notes:

About the unresolvable error

Theoretically, if there is infinite amount of data and a perfect model, the error can be eliminated.

Actually, all models are imperfect, and our data volume is limited.

<!-- Slide number: 45 -->

Variance and Bias

Combinations of variance and bias are as follows:

Low bias & low variance –> Good model

Low bias & high variance

High bias & low variance

High bias & high variance –> Poor model

Ideally, we want a model that can accurately capture the rules in the training data and summarize the invisible data (new data). However, it is usually impossible for the model to complete both tasks at the same time.

Notes:

<!-- Slide number: 46 -->

Model Complexity and Error

As the model complexity increases, the training error decreases.

As the model complexity increases, the test error decreases to a certain point and then increases in the reverse direction, forming a convex curve.

Low bias & high variance

High bias & low variance



Testing error

Error

Model Complexity

Training error

Notes:

<!-- Slide number: 47 -->

Machine Learning Performance Evaluation - Regression

Notes:

<!-- Slide number: 48 -->

Machine Learning Performance Evaluation - Classification (1)

Confusion matrix

Notes:

<!-- Slide number: 49 -->

Machine Learning Performance Evaluation - Classification (2)

Notes:

<!-- Slide number: 50 -->

Example of Machine Learning Performance Evaluation

Notes:

<!-- Slide number: 51 -->

Machine Learning Definition

Machine Learning Types

Machine Learning Process

Other Key Machine Learning Methods

Common Machine Learning Algorithms

Case study

Notes:

<!-- Slide number: 52 -->

Machine Learning Training Method - Gradient Descent (1)

Cost surface



Notes:

<!-- Slide number: 53 -->

Machine Learning Training Method - Gradient Descent (2)

Batch Gradient Descent (BGD) uses the samples (m in total) in all datasets to update the weight parameter based on the gradient value at the current point.

Stochastic Gradient Descent (SGD) randomly selects a sample in a dataset to update the weight parameter based on the gradient value at the current point.

Mini-Batch Gradient Descent (MBGD) combines the features of BGD and SGD and selects the gradients of n samples in a dataset to update the weight parameter.

Notes:

<!-- Slide number: 54 -->

Machine Learning Training Method - Gradient Descent (3)

Comparison of three gradient descent methods

In the SGD, samples selected for each training are stochastic. Such instability causes the loss function to be unstable or even causes reverse displacement when the loss function decreases to the lowest point.

BGD has the highest stability but consumes too many computing resources. MBGD is a method that balances SGD and BGD.

BGD

Uses all training samples for training each time.

SGD

Uses one training sample for training each time.

MBGD

Uses a certain number of training samples for training each time.

Notes:

Reverse displacement is caused by noise data from datasets.

<!-- Slide number: 55 -->

Parameters and Hyperparameters in Models

The model contains not only parameters but also hyperparameters. The purpose is to enable the model to learn the optimal parameters.

Parameters are automatically learned by models.

Hyperparameters are manually set.


Model parameters are "distilled" from data.

Model

Training

Use hyperparameters to control training.

Notes:

<!-- Slide number: 56 -->

Hyperparameters of a Model

Often used in model parameter estimation processes.

Often specified by the practitioner.

Can often be set using heuristics.

Often tuned for a given predictive modeling problem.

Model hyperparameters are external configurations of models.

Common model hyperparameters

Notes:

Parameters are a part of a model that is learned from historical training data and key to machine learning algorithms. They have the following features:

They are required by a model to make predictions.

Their values define the model functions.

They are obtained by data estimation or data learning.

They are often not set manually by the practitioner.

They are often saved as a part of the learned model.

Examples:

Weights in an artificial neural network

Support vectors in a support vector machine

Coefficients in linear regression or logistic regression

<!-- Slide number: 57 -->

Hyperparameter Search Procedure and Method

Procedure for searching hyperparameters

Dividing a dataset into a training set, validation set, and test set.

Optimizing the model parameters using the training set based on the model performance indicators.

Searching for the model hyper-parameters using the validation set based on the model performance indicators.

Perform step 2 and step 3 alternately. Finally, determine the model parameters and hyperparameters and assess the model using the test set.

Search algorithm (step 3)

Grid search

Random search

Heuristic intelligent search

Bayesian search

Notes:

<!-- Slide number: 58 -->

Hyperparameter Searching Method - Grid Search

Grid search attempts to exhaustively search all possible hyperparameter combinations to form a hyperparameter value grid.

In practice, the range of hyperparameter values to search is specified manually.

Grid search is an expensive and time-consuming method.

This method works well when the number of hyperparameters is relatively small. Therefore, it is applicable to generally machine learning algorithms but inapplicable to neural networks (see the deep learning part).

Grid search

5

Hyperparameter 1

4

3

2

1

4

2

0

1

5

3

Hyperparameter 2

Notes:

<!-- Slide number: 59 -->

Hyperparameter Searching Method - Random Search

When the hyperparameter search space is large, random search is better than grid search.

In random search, each setting is sampled from the distribution of possible parameter values, in an attempt to find the best subset of hyperparameters.

Note:

Search is performed within a coarse range, which then will be narrowed based on where the best result appears.

Some hyperparameters are more important than others, and the search deviation will be affected during random search.

Random search

Parameter 1

Parameter 2

Notes:

<!-- Slide number: 60 -->

Cross Validation (1)

Notes:

<!-- Slide number: 61 -->

Cross Validation (2)

Entire dataset

Training set

Test set

Training set

Validation set

Test set

Note: The K value in K-fold cross validation is also a hyperparameter.

Notes:

<!-- Slide number: 62 -->

Machine Learning Definition

Machine Learning Types

Machine Learning Process

Other Key Machine Learning Methods

Common Machine Learning Algorithms

Case study

Notes:

<!-- Slide number: 63 -->

Machine Learning Algorithm Overview

Machine learning

Unsupervised learning

Supervised learning

Classification

Regression

Clustering

Others

Linear regression

Logistic regression

Correlation rule

K-means

Hierarchical clustering

Principal component analysis (PCA)

SVM

SVM

Density-based clustering

Gaussian mixture model (GMM)

Neural network

Neural network

Decision tree

Decision tree

Random forest

Random forest

GBDT

GBDT

KNN

Naive Bayes

Notes:

<!-- Slide number: 64 -->

Linear Regression (1)

Linear regression: a statistical analysis method to determine the quantitative relationships between two or more variables through regression analysis in mathematical statistics.

Linear regression is a type of supervised learning.

Multi-dimensional linear regression

Unary linear regression

Notes:

<!-- Slide number: 65 -->

Linear Regression (2)

Notes:

<!-- Slide number: 66 -->

Linear Regression Extension - Polynomial Regression

Polynomial regression is an extension of linear regression. Generally, the complexity of a dataset exceeds the possibility of fitting by a straight line. That is, obvious underfitting occurs if the original linear regression model is used. The solution is to use polynomial regression.

Comparison between linear regression and polynomial regression

Notes:

<!-- Slide number: 67 -->

Linear Regression and Overfitting Prevention

Notes:

<!-- Slide number: 68 -->

Logistic Regression (1)


Notes:

<!-- Slide number: 69 -->

Logistic Regression (2)

Notes:

After regularization terms are added, logistic regression can also prevent overfitting.

<!-- Slide number: 70 -->

Logistic Regression Extension - Softmax Function (1)

Logistic regression applies only to binary classification problems. For multi-class classification problems, use the Softmax function.

Binary classification problem

Multi-class classification problem

Grape?

Orange?

Male?

Apple?

Female?

Banana?

Notes:

<!-- Slide number: 71 -->

Logistic Regression Extension - Softmax Function (2)

Softmax regression is a generalization of logistic regression that we can use for K-class classification.

The Softmax function is used to map a K-dimensional vector of arbitrary real values to another K-dimensional vector of real values, where each vector element is in the interval (0, 1).

The regression probability function of Softmax is as follows:

Notes:

<!-- Slide number: 72 -->

Logistic Regression Extension - Softmax Function (3)

Softmax assigns a probability to each class in a multi-class problem. These probabilities must add up to 1.

Softmax may produce a form belonging to a particular class. Example:

Category

Probability

Grape?

0.09

Sum of all probabilities:

0.09 + 0.22 + 0.68 + 0.01 =1

Most probably, this picture is an apple.

0.22

Orange?

Apple?

0.68

Banana?

0.01

Notes:

<!-- Slide number: 73 -->

Decision Tree

A decision tree is a tree structure (a binary tree or a non-binary tree). Each non-leaf node represents a test on a feature attribute. Each branch represents the output of a feature attribute in a certain value range, and each leaf node stores a category. To use the decision tree, start from the root node, test the feature attributes of th...