CRISP-DM

Question 1

What are the stages of CRISP-DM

Answer 1

Project Understanding
Data Understanding
Data Preparation
Modeling
Evaluation
Deployment

Question 2

What happens in the Project Understanding Phase?

Answer 2

Problem formulation
Mapping problem to a data mining task
Understanding resources

Question 3

Define Problem Formulation

Answer 3

Defining:
-objectives/deliverables
-criteria for success
-potential benefits/risks
-constrains/assumptions

for a data analysis project.

Question 4

Problem Formulation stage - time used vs importance

Answer 4

Time: 20%
Importance:80%

Question 5

What does ‘mapping problem to data mining task’ mean?

Answer 5

Part of the Project Understanding Phase.
Choosing data analysis method to use: classification, regression, deviation analysis, association analysis.

Question 6

What can be done to reduce communication problems in Project Understanding phase?

Answer 6

For ex:
Cognitive mapping
Rephrasing

Question 7

What are cognitive maps?

Answer 7

A visual representation of a problem that supports domain understanding.

Nodes are domain variables, arrows show direction and type of influence between nodes.

Question 8

What happens in the Data Understanding Phase?

Answer 8

Collecting the data
Using exploratory data analysis to get familiar with data and gain insights.
Evaluating data quality
Selecting interesting subsets that may contain patterns.

Question 9

What kind of goals can a data mining project have?

Answer 9

Type of data mining task (regression, classification, clustering)
Predictive accuracy, model flexibility, interpretability, run-time

Question 10

What types of methods are available for solving different types of problems?

Answer 10

Classification, regression, clustering, association analysis, neural networks.

Question 11

How can different requirements or desirable properties affect our choice of method?

Answer 11

If it needs to be highly interpretable, neural networks wouldn’t be a good fit. Some methods take more computational resources.

Question 12

How can a project go wrong from the very start?

Answer 12

If the data does not support the project goals.

Question 13

What is the main goal of Data Understanding phase?

Question 14

Rule of thumb for Data Understanding?

Answer 14

Never trust any data before some plausibility tests!

Question 15

What types of attributes are there?

Question 16

Can categorial attributes be represented as numerical?

Question 17

What are some typical problems with data quality?

Question 18

What are some data visualization techniques and what can they reveal?

Question 19

What does z-score standardization do?

Question 20

Why might you need z-score standardization?

Question 20

Why might you need z-score standardization?

Question 21

What is dimensionality reduction?

Question 22

Compare dimensionality reduction to feature selection.

Question 23

How to select the number of principal components to use?

Question 24

What are outliers?

Question 25

How can outliers be detected?

Question 26

What are missing values?

Question 27

Are some types of missing values more difficult to deal with than others?

Question 28

When does data preparation happen?

Answer 28

A part of it needs to happen in the data understanding phase, more in the data preparation phase in a more principled manner.

Question 29

What happens in Data Preparation?

Answer 29

Feature extraction
Data quality problems dealt with
Normalization of features.
Dimensionality reduction.

Question 30

What is feature extraction?

Question 31

Why do feature selection?

Question 32

Why can it be beneficial to normalize feature values?

Question 33

Overall problem types

Answer 33

Classification, regression, clustering, association analysis.

Question 34

Model Structures

Answer 34

Nearest Neighbor predictor, linear model, if-else rule statements, decision trees, neural network, clustering.

Question 35

What is the goal of a machine learning algorithm?

Answer 35

To find a good model that fits the problem.

Question 36

How to choose between different models?

Answer 36

For example Cross-validation.

Question 37

What two criteria need to be balanced in learning?

Answer 37

1) Fit to data (low error)
2) Model complexity

Question 38

Define Overfitting

Answer 38

The model is too complex, and may simply model noise in data.

Question 39

Underfitting

Answer 39

The model is too simple and may not capture the concept.

Question 40

The more data your have, the more –

Answer 40

– complex models can you afford to use.

Question 41

Many theoretical approaches to defining what we mean by
complex:

Answer 41

Regularization theory, minimum description length
principle, Bayesian priors.

Question 42

What does a hyperparameter do?

Question 43

Examples of hyperparameters?

Answer 43

The max degree of polynomials on the above regression

The number of neighbors in the k-nearest neighbor method

The regularization parameter value for many methods
(regularized linear regression models, support vector
machines,…)
Depth of decision tree
The number of layers and the number of neurons per hidden
layer with neural networks
You might also be comparing models produced by different
algorithms altogether