Explain vs Predict, Data Preprocessing and Predictive Modeling

Question 1

What is the goal of:

Explanatory modeling?

Answer 1

Explanatory modeling is a theory based approach to test causal hypotheses.

Question 2

What is the goal of:

Predictive modeling?

Answer 2

Predictive modeling is using data science methods to make predictions.

Question 3

How is explanatory modeling evaluated?

Answer 3

Explanatory modeling is evaluated by the strength of relationship in statistical model.

Question 3

How is explanatory modeling evaluated?

Answer 3

Explanatory modeling is evaluated by the strength of relationship in statistical model

Question 4

How is predictive power evaluated?

Answer 4

Predictive power is evaluated by ability of the model to accurately predict new observations.

Question 5

How does data preparation for ‘explanatory modeling’ and ‘predictive modeling’ differ concerning ‘missing values’?

Answer 5

Explanatory modeling: throw away

Predictive modeling: throwing away is not an option if we need to make a prediction for these, since it can even be predictive information.

Question 6

How does data preparation for ‘explanatory modeling’ and ‘predictive modeling’ differ concerning ‘data partitioning’?

Answer 6

Predictive: test set (we’ll come back to this) crucial: How well can we predict on new, unseen data instances?

Explanatory: much less common

Question 7

How does data preparation for ‘explanatory modeling’ and ‘predictive modeling’ differ concerning ‘the choice of variables’?

Answer 7

Explanatory: operationalization of constructs

Predictive: more broad, but it must be available at time of prediction

Question 8

What is the difference in methods for Explanatory vs Predictive Modeling?

Answer 8

Explanatory: interpretable, statistical methods

Predictive: accurate data mining methods (neural networks, random forests, etc. But also logistic regression)

Question 9

What is the difference in validation of the results of Explanatory vs Predictive Modeling?

Answer 9

Explanatory modeling: model fit, R-squared

Predictive modeling: generalization, (test)-accuracy, etc.

Question 10

What is wrong with this statement:

“We checked for multicollinearity of the input variables before their use for prediction.”

Answer 10

For statistical exercices the coefficient estimates of the multiple regression may change erratically.
Multicollinearity will not affect the ability of the model to predict.

Question 11

What is wrong with this statement in data mining:

“Variables income and car_brand were very explanatory for the model.”

Answer 11

Variables are not explanatory but predictive

Question 12

What is:
Sampling?

Why would you do it?

Answer 12

Sampling is the act, process or technique of selecting a suitable sample or a representative part of the population for the purpose of determining parameters or characteristics of the whole population.

There are various reasons to do so: 
Economic advantage: less costs
Time factor: less time, quickly
Large Populations and partly accessible populations
Computation Power Required

Question 13

What is:

Descretization?

Question 14

What is:

Normalization?

Question 15

What are:

Missing Values?

Question 16

What are:

Outliers?

Question 17

What is:

Encoding?

Question 18

What is:

Stratified sampling?

Answer 18

Stratified sampling is the act of giving the same distribution (of some variable/s) in the sample as it is in the population.

Question 19

What is:

kNN?

Answer 19

kNN, or short for K-Nearest Neighbor, is a technique that classifies an instance according to the most similar observation in (for example) Euclidean distance. So the predicted class is the same as the most similar observation.

Question 20

What is a problems of kNN?

Answer 20

A model could potentially assume order in a nominal variable, which is not there, Euclidian distances won’t work for the kNN-technique.

Solution for nominal variable: dummy encoding

Question 21

What is another problem is kNN?

Answer 21

A model could have multiple variables of a higher order (income vs amount of cars in a household), so these have a higher weight, without being more important in the distance.

Solution: Thermometer encoding or standardization