Depending on your downstream system, you may need to convert your data into numeric values of the expected form or to standardize the distribution of numeric values. This section summarizes some common statistical transformations that can be applied to columnar data to prepare it for use in downstream analytic systems. |
You can scale the values within a column using either of the following techniques.
Zero mean and unit variance scaling renders the values in the set to fit a normal distribution with a mean of 0
and a variance of 1
. This technique is a common standard for normalizing values into a normal distribution for statistical purposes.
In the following example, the values in the POS_Sales
column have been normalized to average 0
, variance 1
.
Remove mean: When selected, the existing mean (average) of the values is used as the center of the distribution curve.
NOTE: Re-centering sparse data by removing the mean may remove sparseness. |
Scale to unit variance: When selected, the range of values are scaled such that their variance is 1
. When deselected, the existing variance is maintained.
NOTE: Scaling to unit variance may not work well for managing outliers. Some additional techniques for managing outliers are outlined below. |
You can scale column values fitting between a specified minimum and maximum value. This technique is useful for distributions with very small standard deviation values and for preserving 0 values in sparse data.
The following example scales the TestScores
column to a range of 0
and 1
, inclusive.
You can use several techniques for identifying statistical outliers in your dataset and managing them as needed.
Suppose you need to remove the outliers from a column. Assuming a normal bell distribution of values, you can use the following formula to calculate the number of standard deviations a column value is from the column mean (average). In this case, the source column is POS_Sales
.
The new stdevs_POS_Sales
column now contains the number of standard deviations from the mean for the corresponding value in POS_Sales
. You can use the following transformation to remove the rows that contain outlier values for this column.
Tip: An easier way to select these outlier values is to select the range of values in the |
In the following transformation, all rows that contain a value in POS_Sales
that is greater than four standard deviations from the mean are deleted:
You can also remove the effects of outliers be setting their value to the mean (average), which preserves the data in other columns in the row.
You can modify your data to fit into bins of equal or custom size. For example, the lowest values in your range would be marked in the 0
bin, with larger values being marked with larger bin numbers.
You can bin numeric values into bins of equal size. Suppose your column contains numeric values 0-1000
. You can bin values into equal ranges of 100
by creating 10
bins.
You can also create custom bins. In the following example, the TestScores
column is binned into the following bins. In a later step, these bins are mapped to grades:
Bins | Bin Range | Bin Number | Grade |
---|---|---|---|
59 | 0-59 | 0 | F |
69 | 60-69 | 1 | D |
79 | 70-79 | 2 | C |
89 | 80-89 | 3 | B |
90+ | 4 | A | |
(no value) | I |
First, you bin values into the bin numbers listed above:
You can then use the following transformation to assign letters in the Grades
column:
One-hot encoding refers to distributing the listed values in a column into individual columns. Within each row of each individual column is a 0
or a 1
, depending on whether the value represented by the column appears in the corresponding source column. The source column is untouched. This method of encoding allows for easier consumption of data in target systems.
Tip: This transformation is particularly useful for columns containing a limited set of enumerated values. |
In the following example, the values in the BrandName
column are distributed into separate columns of binary values, with a maximum limit of 50
new columns.
NOTE: Be careful applying this to a column containing a wide variety of values, such as Decimal values. Your dataset can expand significantly in size. Use the max columns setting to constrain the upper limit on dataset expansion. |
Tip: If needed, you can rename the columns to prepend the names with a reference to the source column. |