How to Transform Numerical to Categorical Data: Suitable For Highly Skewed or Non-Standard Distribution

Question 1

WHAT DO DISCRETIZATION TRANSFORMS DO? P303

Answer 1

Discretization transforms are a technique for transforming numerical input or output variables to have:
Discrete ordinal labels
Different data distribution

Question 2

WHICH LIBRARY DO WE USE FOR CHANGING THE STRUCTURE AND DISTRIBUTION OF NUMERIC VARIABLES TO CATEGORICAL TO IMPROVE THE PERFORMANCE OF PREDICTIVE MODELS? P303

Answer 2

KBinsDiscretizer

Question 3

WHAT ARE 3 COMMON METHODS WE CAN USE FOR GROUPING VALUES INTO K DISCRETE BINS? P305

Answer 3

ˆ Uniform: All bins in each feature have identical widths.
ˆ Quantile: All bins in each feature have the approximately same number of points.
ˆ Kmeans: Clusters are identified and examples are assigned to each group.

Question 4

WHAT ARE THE VALUES WE CAN USE FOR ‘STRATEGY’ PARAMETER OF KBINSDISCRETIZER? P305

Answer 4

Uniform, quantile, kmeans.

Question 5

WHAT DOES N_BINS PARAMETER MEAN, IN KBINSDISCRETIZER BASED ON WHAT IS IT SET? P305

Answer 5

It controls the number of bins that will be created.
It must be set based on the choice of strategy.

Question 6

HOW IS THE N_BINS CHOSEN FOR DIFFERENT STRATEGIES IN KBINSDISCRETIZER? P305

Answer 6

Uniform: flexible
Quantile: n_bins less than the number of observations or sensible percentile.
K-means: a value for the number of clusters that can be reasonably found

Question 7

WHAT DOES ENCODE ARGUMENT CONTROLS IN KBINSDISCRETIZER? P305

Answer 7

It controls whether the transform will map each values to an integer value by setting “ordinal” or a one hot encoding “onehot”.

Question 8

WHICH METHOD OF ENCODING IS PREFERRED IN KBINSDISCRETIZER? P305

Answer 8

Ordinal

Question 9

WHEN DO WE USE ONEHOT, FOR ENCODE PARAMETER IN KBINSDISCRETIZER? P305

Answer 9

For example in the case of k-means clustering strategy

Question 10

WHAT SORT OF RELATIONSHIPS CAN A MODEL LEARN WHEN WE USE ONEHOT ENCODING? P305

Answer 10

Non-ordinal relationships

Question 11

WHICH RANGE IS SUITABLE FOR K-MEANS STRATEGY’S N_BINS IN KBINSDISCRETIZER? P312

Answer 11

3-5, unless the empirical distribution of the variable is complex.

Question 12

DOES UNIFORM DISCRETIZATION TRANSFORM CHANGE THE PROBABILITY DISTRIBUTION? P309

Answer 12

No, Me: because the bin widths are all the same, it’s like creating a type of histogram

Question 13

WHAT RANGE OF NUMBER IS SUITABLE FOR N_BINS OF QUANTILE STRATEGY IN KBINSDISCRETIZATION TRANSFORM? P315

Answer 13

5-10, unless there are a large number of observations or a complex empirical distribution, this number should be kept small.