Parallel & Scalable Data Analysis

Introduction to Machine Learning Algorithms

Dr. – Ing. Morris Riedel

Adjunct Associated Professor
School of Engineering and Natural Sciences, University of Iceland
Research Group Leader, Juelich Supercomputing Centre, Germany

LECTURE 1

Machine Learning Fundamentals

November 23th, 2017
Ghent, Belgium
 Outline

Lecture 1 – Machine Learning Fundamentals 2 / 60

 Outline of the Course

1. Machine Learning Fundamentals

2. Unsupervised Clustering and Applications

3. Supervised Classification and Applications

4. Classification Challenges and Solutions

5. Regularization and Support Vector Machines

6. Validation and Parallelization Benefits

Lecture 1 – Machine Learning Fundamentals 3 / 60

 Outline

 Machine Learning Basics

 Motivation
 Methods Overview
 Simple Application Example
 Perceptron Learning Model
 Decision Boundary & Linear Seperability

 Learning from Data

 Systematic Process to Support Learning
 Predictive and Descriptive Tasks
 Different Learning Approaches
 Terminologies
 Model Evaluation with Testing

Lecture 1 – Machine Learning Fundamentals 4 / 60

 Machine Learning Basics

Lecture 1 – Machine Learning Fundamentals 5 / 60

 Motivation

 Rapid advances in data collection and

storage technologies in the last decade
 Extracting useful information is a challenge
considering ever increasing massive datasets
 Traditional data analysis techniques cannot be
used in growing cases (e.g. memory limits)

 Machine learning / Data Mining is a technology that blends traditional data analysis
methods with sophisticated algorithms for processing large volumes of data
 Machine Learning / Data Mining is the process of automatically discovering useful
information in large data repositories ideally following a systematic process
modified from [1] Introduction to Data Mining

 Machine Learning & Statistical Data Mining

 Traditional statistical approaches are still very useful to consider
 E.g. in order to reduce large quantities of data to most expressive datasets
Lecture 1 – Machine Learning Fundamentals 6 / 60
 Machine Learning Prerequisites

1. Some pattern exists

2. No exact mathematical formula
3. Data exists

 Idea ‘Learning from Data‘

Data Applied
shared with a wide variety Mining Statistics
of other disciplines Data
Science
 E.g. signal processing,
data mining, etc.
 Challenge: Data is often complex
Machine
Learning
 Machine learning is a very broad subject and goes from
very abstract theory to extreme practice (‘rules of thumb’)

Lecture 1 – Machine Learning Fundamentals 7 / 60

 Examples of Real Data Collections

 Data collection of the earth and environmental science domain

 Different from the known ‘UCI machine learning repository examples‘

(real science datasets examples) (examples for learning & comparisons)

[2] PANGAEA data collection [3] UCI Machine Learning Repository

Lecture 1 – Machine Learning Fundamentals 8 / 60

 Methods Overview
 Machine learning methods can be roughly categorized in classification, clustering, or regression
augmented with various techniques for data exploration, selection, or reduction

Classification Clustering Regression

 Groups of data exist  No groups of data exist  Identify a line with

 New data classified  Create groups from a certain slope
to existing groups data close to each other describing the data
 The concrete focus of this course is classification using one specific technique out of many others
Lecture 1 – Machine Learning Fundamentals 9 / 60
 Simple Application Example: Classification of a Flower

(1) Problem Understanding Phase (what type of flower is this?)

(flowers of type ‘IRIS Setosa‘)

 Groups of data exist

 New data classified
to existing groups

[4] Image sources: Species Iris Group of (flowers of type ‘IRIS Virginica‘)
North America Database, [Link]
Lecture 1 – Machine Learning Fundamentals 10 / 60
 The Learning Problem in the Example

(flowers of type ‘IRIS Setosa‘) (flowers of type ‘IRIS Virginica‘)

[4] Image sources: Species Iris Group of North America Database, [Link]

Learning problem: A prediction task

 Determine whether a new Iris flower
sample is a “Setosa” or “Virginica”
 Binary (two class) classification problem
 What attributes about the data help?
(what type of flower is this?)
Lecture 1 – Machine Learning Fundamentals 11 / 60
 Feasibility of Machine Learning in this Example

1. Some pattern exists:

 Believe in a ‘pattern with ‘petal length‘ &
‘petal width‘ somehow influence the type
2. No exact mathematical formula
 To the best of our knowledge there is no
precise formula for this problem
3. Data exists
 Data collection from UCI Dataset „Iris“
 150 labelled samples (aka ‘data points‘) [5] Image source: Wikipedia, Sepal

 Balanced: 50 samples / class  sepal length in cm

(four data attributes for each  sepal width in cm
(2) Data Understanding Phase
sample in the dataset)  petal length in cm
 petal width in cm
 class: Iris Setosa, or
[6] UCI Machine Learning (one class label for each
Repository Iris Dataset sample in the dataset)
Iris Versicolour, or
Lecture 1 – Machine Learning Fundamentals
Iris Virginica 12 / 60
 Exercises

Lecture 1 – Machine Learning Fundamentals 13 / 60

 Understanding the Data – Check Metadata

 First: Check metadata if available (metadata is not always available in practice)

 Example: Downloaded [Link] includes metadata about data

(Subject, title, or context)

(author, source, or creator)

… (number of samples, instances)

(attribute information)

(detailed attribute
information)

(detailed attribute
information)

[6] UCI Machine Learning Repository Iris Dataset

Lecture 1 – Machine Learning Fundamentals 14 / 60
 Understanding the Data – Check Table View

 Second: Check table view of the dataset with some samples

 E.g. Using a GUI like ‘Rattle‘ (library of R), or Excel in Windows, etc.
 E.g. Check the first row if there is header information or if is a sample

(careful first sample taken as header,

resulting in only 149 data samples)

 sepal length in cm
(four data attributes for each  sepal width in cm
sample in the dataset)  petal length in cm
 petal width in cm
 class: Iris Setosa, or
(one class label for each
sample in the dataset)
Iris Versicolour, or
Iris Virginica

[7] Rattle Library for R

Lecture 1 – Machine Learning Fundamentals 15 / 60
 Preparing the Data – Corrected Header

(3) Data Preparation Phase

(correct header information, resulting in 150 data samples)

(correcting the header is not always necessary,

or can be automated, e.g. in Rattle)

Lecture 1 – Machine Learning Fundamentals 16 / 60

 Preparing the Data – Remove Third Class Samples

 Data preparation means to prepare our data for our problem

 In practice the whole dataset is rarely needed to solve one problem
 E.g. apply several sampling strategies (but be aware of class balance)
 Recall: Our learning problem
 Determine whether a new Iris flower sample is a “Setosa” or “Virginica”
 Binary (two class) classification problem : ‘Setosa‘ or ‘Virginica‘

(three class problem with (wo class problem with

N = 150 samples including N = 100 samples excluding
Iris Versicolour) Iris Versicolour)

(remove Versicolour class (export or save dataset

samples from dataset) to [Link])

Lecture 1 – Machine Learning Fundamentals 17 / 60

 Preparing the Data – Feature Selection Process

 Data preparation means to prepare our data for our problem

 In practice the whole dataset is rarely needed to solve one problem
 E.g. perform feature selection (aka remove not needed attributes)
 Recall: Our believed pattern in the data
 A ‘pattern with ‘petal length‘ & ‘petal width‘ somehow influence the type

 sepal length in cm  petal length in cm

 sepal width in cm  petal width in cm
 petal length in cm  class: Iris Setosa, or
 petal width in cm Iris Versicolour, or
 class: Iris Setosa, or Iris Virginica
Iris Versicolour, or (export or save dataset
Iris Virginica to [Link])

(N = 100 samples with 4 attributes and 1 class label) (N = 100 samples with 2 attributes and 1 class label)
Lecture 1 – Machine Learning Fundamentals 18 / 60
 Exercises

Lecture 1 – Machine Learning Fundamentals 19 / 60

 Check Preparation Phase: Plotting the Data
Dataset
3

petal width (in cm)

2.5 (attributes with d=2)

2 (x1 is petal length,

x2 is petal width)

1.5
Dataset

(Recall: we believed in a ‘pattern‘ with ‘petal length‘

1
& ‘petal width‘ somehow influence the flower type)
(N = 100 samples)

0.5
(what about the class labels?)

0
0 1 2 3 4 5 6 7 8
petal length (in cm)

Lecture 1 – Machine Learning Fundamentals 20 / 60

 Check Preparation Phase: Class Labels

petal width (in cm)

2.5

1.5 Iris-setosa
Iris-virginica

1
(N = 100 samples)

0.5

(still no machine learning so far)

0 petal length (in cm)

0 1 2 3 4 5 6 7 8

Lecture 1 – Machine Learning Fundamentals 21 / 60

 Linearly Seperable Data & Linear Decision Boundary

3
(4) Modelling Phase  The data is
linearly seperable
petal width (in cm) (rarely in practice)
2.5  A line becomes a
decision boundary
to determine if a
2 new data point is
class red/green
1.5
? Iris-setosa
Iris-virginica

1
(N = 100 samples)

0.5

(decision boundary)

0 petal length (in cm)

0 1 2 3 4 5 6 7 8

Lecture 1 – Machine Learning Fundamentals 22 / 60

 Separating Line & Mathematical Notation

 Data exploration results

 A line can be crafted between the classes since linearly seperable data
 All the data points representing Iris-setosa will be below the line
 All the data points representing Iris-virginica will be above the line
 More formal mathematical notation
 Input: (attributes of flowers)

 Output: class +1 (Iris-virginica) or class -1 (Iris-setosa)

(decision boundary)

Iris-virginica if
(wi and threshold are
still unknown to us)
Iris-setosa if

(compact notation)

Lecture 1 – Machine Learning Fundamentals 23 / 60

 Separating Line & ‘Decision Space‘ Example

(decision boundary)

(equation of a line)

(all points Xi on this line

have to satisfy this equation)

modified from [13] An Introduction to Statistical Learning

Lecture 1 – Machine Learning Fundamentals 24 / 60

 A Simple Linear Learning Model – The Perceptron

 Human analogy in learning [8] F. Rosenblatt, 1957

 Human brain consists of nerve cells called neurons

 Human brain learns by changing the strength of neuron connections (wi)
upon repeated stimulation by the same impulse (aka a ‘training phase‘)
 Training a perceptron model means adapting the weights wi
 Done until they fit input-output relationships of the given ‘training data‘

(training data)

(modelled as
bias term)

(activation
function,
d
+1 or -1) (representing the threshold)
(the signal) (dimension of features)

Lecture 1 – Machine Learning Fundamentals 25 / 60

 Perceptron – Example of a Boolean Function

(training data)

(training phase)

(trained perceptron model)

 Output node interpretation

 More than just the weighted sum of the inputs – threshold (aka bias)
 Activation function sign (weighted sum): takes sign of the resulting sum
(e.g. consider sample #3,
sum is positive (0.2)  +1)
(e.g. consider sample #6,
sum is negative (-0.1)  -1)

Lecture 1 – Machine Learning Fundamentals 26 / 60

 Summary Perceptron & Hypothesis Set h(x)

 When: Solving a linear classification problem [8] F. Rosenblatt, 1957

 Goal: learn a simple value (+1/-1) above/below a certain threshold

 Class label renamed: Iris-setosa = -1 and Iris-virginica = +1
 Input: (attributes in one dataset)

 Linear formula (take attributes and give them different weights – think of ‘impact of the attribute‘)
 All learned formulas are different hypothesis for the given problem
(parameters that define
one hypothesis vs. another)
(each green space and
blue space are regions
of the same class label (red parameters correspond
determined by sign to the redline in graphics)
function)
(but question remains: how do
we actually learn wi and threshold?)

Lecture 1 – Machine Learning Fundamentals 27 / 60

Perceptron Learning Algorithm – Understanding Vector W

 When: If we believe there is a linear pattern to be detected

 Assumption: Linearly seperable data (lets the algorithm converge)
 Decision boundary: perpendicular vector wi fixes orientation of the line

wi (vector notation, using T = transpose)

(points on the decision

boundary satisfy this equation)

 Possible via simplifications since

we also need to learn the threshold:

(equivalent dotproduct notation)

[9] Rosenblatt, 1958 (all notations are equivalent and result is a scalar from which we derive the sign)
Lecture 1 – Machine Learning Fundamentals 28 / 60
Understanding the Dot Product – Example & Interpretation

 ‘Dot product‘
 Given two vectors
 Multiplying corresponding components of the vector
(our example)
 Then adding the resulting products
 Simple example: (a scalar!)

 Interesting: Dot product of two vectors is a scalar

 ‘Projection capabilities of Dot product‘ (simplified)

 Orthogonal projection of vector in the direction of vector
(projection)

 Normalize using length of vector

Lecture 1 – Machine Learning Fundamentals 29 / 60

 Perceptron Learning Algorithm – Learning Step

 Iterative Method using (labelled) training data

(one point at a time is picked)
1. Pick one misclassified
y = +1 w + yx
training point where:
(a) w x

2. Update the weight vector: (a) adding a vector or

(b) subtracting a vector

(yn is either +1 or -1)

y = -1
w
 Terminates when there are
no misclassified points x
(b)
(converges only with linearly seperable data)
w – yx

Lecture 1 – Machine Learning Fundamentals 30 / 60

 [Video] Perceptron Learning Algorithm

[10] PLA Video

Lecture 1 – Machine Learning Fundamentals 31 / 60

 Exercises

Lecture 1 – Machine Learning Fundamentals 32 / 60

 Learning from Data

Lecture 1 – Machine Learning Fundamentals 33 / 60

 Systematic Process to Support Learning From Data

 Systematic data analysis guided by a ‘standard process‘

 Cross-Industry Standard Process for Data Mining (CRISP-DM)

 A data mining project is

guided by these six phases:
(1) Problem Understanding;
(2) Data Understanding;
(3) Data Preparation;
(4) Modeling; (learning
(5) Evaluation; takes place)
(6) Deployment

 Lessons Learned from Practice

 Go back and forth between [11] C. Shearer, CRISP-DM model,
the different six phases Journal Data Warehousing, 5:13

 A more detailed description of all six CRISP-DM phases is in the appendix of the slideset
Lecture 1 – Machine Learning Fundamentals 34 / 60
 Machine Learning & Data Mining Tasks in Applications

 Machine learning tasks can be divided into two major categories: Predictive and Descriptive Tasks
[1] Introduction to Data Mining
 Predictive Tasks
 Predicts the value of an attribute based on values of other attributes
 Target/dependent variable: attribute to be predicted
 Explanatory/independent variables: attributed used for making predictions
 E.g. predicting the species of a flower based on characteristics of a flower

 Descriptive Tasks
 Derive patterns that summarize the underlying relationships in the data
 Patterns here can refer to correlations, trends, trajectories, anomalies
 Often exploratory in nature and frequently require postprocessing
 E.g. credit card fraud detection with unusual transactions for owners

Lecture 1 – Machine Learning Fundamentals 35 / 60

Predicting Task: Obtain Class of a new Flower ‘Data Point‘

3
(4) Modelling Phase

petal width (in cm)

2.5

1.5
? Iris-setosa
Iris-virginica

1
(N = 100 samples)

0.5

(decision boundary)

0 petal length (in cm)

0 1 2 3 4 5 6 7 8

[4] Image sources: Species Iris Group of North America Database, [Link]
Lecture 1 – Machine Learning Fundamentals 36 / 60
 What means Learning?

 The basic meaning of learning is ‘to use a set of observations to uncover an underlying process‘
 The three different learning approaches are supervised, unsupervised, and reinforcement learning

 Supervised Learning
 Majority of methods follow this approach in this course
 Example: credit card approval based on previous customer applications
 Unsupervised Learning
 Often applied before other learning  higher level data representation
 Example: Coin recognition in vending machine based on weight and size
 Reinforcement Learning
 Typical ‘human way‘ of learning
 Example: Toddler tries to touch a hot cup of tea (again and again)

Lecture 1 – Machine Learning Fundamentals 37 / 60

 Learning Approaches – Supervised Learning

 Each observation of the predictor measurement(s)

has an associated response measurement:
 Input
 Output
 Data
 Goal: Fit a model that relates the response to the predictors
 Prediction: Aims of accurately predicting the response for future
observations
 Inference: Aims to better understanding the relationship between the
response and the predictors

 Supervised learning approaches fits a model that related the response to the predictors
 Supervised learning approaches are used in classification algorithms such as SVMs
 Supervised learning works with data = [input, correct output]
[13] An Introduction to Statistical Learning
Lecture 1 – Machine Learning Fundamentals 38 / 60
 Learning Approaches – Supervised Learning Example

3
 The labels guide
our learning
petal width (in cm)

process like a
2.5 ‘supervisor‘ is
helping us

1.5
? Iris-setosa
Iris-virginica

1
(N = 100 samples)

0.5

(decision boundary)

0 petal length (in cm)

0 1 2 3 4 5 6 7 8

 Lecture 2 provides details on the supervised learning approach using classification

Lecture 1 – Machine Learning Fundamentals 39 / 60
 Learning Approaches – Unsupervised Learning

 Each observation of the predictor measurement(s)

has no associated response measurement:
 Input
 No output
 Data
 Goal: Seek to understand relationships between the observations
 Clustering analysis: check whether the observations fall into distinct groups
 Challenges
 No response/output that could supervise our data analysis
 Clustering groups that overlap might be hardly recognized as distinct group
 Unsupervised learning approaches seek to understand relationships between the observations
 Unsupervised learning approaches are used in clustering algorithms such as k-means, etc.
 Unupervised learning works with data = [input, ---]
[13] An Introduction to Statistical Learning
Lecture 1 – Machine Learning Fundamentals 40 / 60
 Learning Approaches – Unsupervised Learning Example

 Practice: The number of clusters

can be ambiguities

[13] An Introduction to Statistical Learning

 Lecture 2 offers more details about unsupervised learning using clustering algorithms in practice
Lecture 1 – Machine Learning Fundamentals 41 / 60
 Learning Approaches – Reinforcement Learning

 Each observation of the predictor measurement(s)

has some associated response measurement:
 Input
 Some output & grade of the output
 Data
 Goal: Learn through iterations
 Guided by output grade: check learning and compare with grade
 Challenge:
 Iterations may require lots of CPU time (e.g. backgammon playing rounds)
 (Rarely tackled in this course, just for the sake of completion)
 Reinforcement learning approaches learn through iterations using the grading output as guide
 Reinforcement learning approaches are used in playing game algorithms (e.g backgammon)
 Unupervised learning works with data = [input, some output, grade for this output]
[13] An Introduction to Statistical Learning
Lecture 1 – Machine Learning Fundamentals 42 / 60
Summary Terminologies & Different Dataset Elements

 Target Function
 Ideal function that ‘explains‘ the data we want to learn
 Labelled Dataset (samples)
 ‘in-sample‘ data given to us:
 Learning vs. Memorizing
 The goal is to create a system that works well ‘out of sample‘
 In other words we want to classify ‘future data‘ (ouf of sample) correct
 Dataset Part One: Training set (4) Modelling Phase
 Used for training a machine learning algorithms
 Result after using a training set: a trained system
 Dataset Part Two: Test set (5) Evaluation Phase

 Used for testing whether the trained system might work well
 Result after using a test set: accuracy of the trained model
Lecture 1 – Machine Learning Fundamentals 43 / 60
 Model Evaluation – Training and Testing Phases

 Different Phases in Learning

(4) Modelling Phase
 Training phase is a hypothesis search
 Testing phase checks if we are on right track (5) Evaluation Phase
(once the hypothesis clear) (e.g. student exam training on examples to
get Ein ‚down‘, then test via exam)
 Work on ‘training examples‘
 Create two disjoint datasets
‘training set’ ‘test set’
 One used for training only
Training Examples
(aka training set)
 Another used for testing only (historical records, groundtruth data, examples)
(aka test set)
 Exact seperation is rule of thumb per use case (e.g. 10 % training, 90% test)
 Practice: If you get a dataset take immediately test data away
(‘throw it into the corner and forget about it during modelling‘)
 Reasoning: Once we learned from training data it has an ‘optimistic bias‘

Lecture 1 – Machine Learning Fundamentals 44 / 60

 Model Evaluation – Testing Phase & Confusion Matrix

 Model is fixed (5) Evaluation Phase

 Model is just used with the testset
 Parameter wi are set and we have a linear decision function
 Evaluation of model performance
 Counts of test records that are incorrectly predicted
 Counts of test records that are correctly predicted
 E.g. create confusion matrix for a two class problem
Counting per sample Predicted Class
Class = 1 Class = 0
Actual Class = 1 f11 f10
Class Class = 0 f01 f00
(serves as a basis for further performance metrics usually used)

Lecture 1 – Machine Learning Fundamentals 45 / 60

Model Evaluation – Testing Phase & Performance Metrics

Counting per sample Predicted Class (5) Evaluation Phase

Class = 1 Class = 0
Actual Class = 1 f11 f10 (100% accuracy in learning often
points to problems using machine
Class Class = 0 f01 f00 learning methos in practice)

 Accuracy (usually in %)

 Error rate

 If model evaluation is satisfactory: (6) Deployment Phase

Lecture 1 – Machine Learning Fundamentals 46 / 60

 Exercises

Lecture 1 – Machine Learning Fundamentals 47 / 60

 [Video] European Plate Observing System

[14] EPOS Data Community Services, YouTube

Lecture 1 – Machine Learning Fundamentals 48 / 60

 Lecture Bibliography

Lecture 1 – Machine Learning Fundamentals 49 / 60

 Lecture Bibliography (1)
 [1] Introduction to Data Mining, Pang-Ning Tan, Michael Steinbach, Vipin Kumar, Addison Wesley, ISBN
0321321367, English, ~769 pages, 2005
 [2] PANGAEA Data Collection, Data Publisher for Earth & Environmental Science,
Online: [Link]
 [3] UCI Machine Learning Repository,
Online: [Link]
 [4] Species Iris Group of North America Database,
Online: [Link]
 [5] UCI Machine Learning Repository Iris Dataset,
Online: [Link]
 [6] Wikipedia ‘Sepal‘,
Online: [Link]
 [7] Rattle Library for R,
Online: [Link]
 [8] F. Rosenblatt, ‘The Perceptron--a perceiving and recognizing automaton’,
Report 85-460-1, Cornell Aeronautical Laboratory, 1957
 [9] Rosenblatt,’The Perceptron: A probabilistic model for information storage and orgainzation in the brain’,
Psychological Review 65(6), pp. 386-408, 1958
 [10] PLA Algorithm, YouTube Video, Online:
 [11] C. Shearer, CRISP-DM model, Journal Data Warehousing, 5:13
 [12] Pete Chapman, ‘CRISP-DM User Guide’, 1999,
Online: [Link]
Lecture 1 – Machine Learning Fundamentals 50 / 60
 Lecture Bibliography (2)
 [13] An Introduction to Statistical Learning with Applications in R, Online:
[Link]
 [14] EPOS - European Plate Observing System -- Community Services, YouTube Video,
Online: [Link]

Lecture 1 – Machine Learning Fundamentals 51 / 60

 Appendix

Lecture 1 – Machine Learning Fundamentals 52 / 60

 Summary: Systematic Process

 Systematic data analysis guided by a ‘standard process‘

 Cross-Industry Standard Process for Data Mining (CRISP-DM)

 A data mining project is

guided by these six phases:
(1) Problem Understanding;
(2) Data Understanding;
(3) Data Preparation;
(4) Modeling;
(5) Evaluation;
(6) Deployment

 Lessons Learned from Practice

 Go back and forth between [11] C. Shearer, CRISP-DM model,
the different six phases Journal Data Warehousing, 5:13

Lecture 1 – Machine Learning Fundamentals 53 / 60

 1 – Problem (Business) Understanding
 The Business Understanding phase consists of four distinct tasks: (A) Determine Business
Objectives; (B) Situation Assessment; (C) Determine Data Mining Goal; (D) Produce Project Plan

 Task A – Determine Business Objectives [12] CRISP-DM User Guide

 Background, Business Objectives, Business Success Criteria

 Task B – Situation Assessment
 Inventory of Resources, Requirements, Assumptions, and Contraints
 Risks and Contingencies, Terminology, Costs & Benefits
 Task C – Determine Data Mining Goal
 Data Mining Goals and Success Criteria
 Task D – Produce Project Plan
 Project Plan
 Initial Assessment of Tools & Techniques

Lecture 1 – Machine Learning Fundamentals 54 / 60

 2 – Data Understanding
 The Data Understanding phase consists of four distinct tasks:
(A) Collect Initial Data; (B) Describe Data; (C) Explore Data; (D) Verify Data Quality

 Task A – Collect Initial Data [12] CRISP-DM User Guide

 Initial Data Collection Report

 Task B – Describe Data
 Data Description Report
 Task C – Explore Data
 Data Exploration Report
 Task D – Verify Data Quality
 Data Quality Report

Lecture 1 – Machine Learning Fundamentals 55 / 60

 3 – Data Preparation
 The Data Preparation phase consists of six distinct tasks: (A) Data Set; (B) Select Data;
(C) Clean Data; (D) Construct Data; (E) Integrate Data; (F) Format Data

 Task A – Data Set [12] CRISP-DM User Guide

 Data set description

 Task B – Select Data
 Rationale for inclusion / exclusion
 Task C – Clean Data
 Data cleaning report
 Task D – Construct Data
 Derived attributes, generated records
 Task E – Integrate Data
 Merged data
 Task F – Format Data
 Reformatted data
Lecture 1 – Machine Learning Fundamentals 56 / 60
 4 – Modeling
 The Data Preparation phase consists of four distinct tasks: (A) Select Modeling
Technique; (B) Generate Test Design; (C) Build Model; (D) Assess Model;

 Task A – Select Modeling Technique [12] CRISP-DM User Guide

 Modeling assumption, modeling technique

 Task B – Generate Test Design
 Test design
 Task C – Build Model
 Parameter settings, models, model description
 Task D – Assess Model
 Model assessment, revised parameter settings

Lecture 1 – Machine Learning Fundamentals 57 / 60

 5 – Evaluation
 The Data Preparation phase consists of three distinct tasks: (A) Evaluate Results;
(B) Review Process; (C) Determine Next Steps

 Task A – Evaluate Results [12] CRISP-DM User Guide

 Assessment of data mining results w.r.t. business success criteria

 List approved models
 Task B – Review Process
 Review of Process
 Task C – Determine Next Steps
 List of possible actions, decision

Lecture 1 – Machine Learning Fundamentals 58 / 60

 6 – Deployment
 The Data Preparation phase consists of three distinct tasks: (A) Plan Deployment;
(B) Plan Monitoring and Maintenance; (C) Produce Final Report; (D) Review Project

 Task A – Plan Deployment [12] CRISP-DM User Guide

 Establish a deployment plan

 Task B – Plan Monitoring and Maintenance
 Create a monitoring and maintenance plan
 Task C – Product Final Report
 Create final report and provide final presentation
 Task D – Review Project
 Document experience, provide documentation

Lecture 1 – Machine Learning Fundamentals 59 / 60

Lecture 1 – Machine Learning Fundamentals 60 / 60