Page tree

Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.
Comment: Published by Scroll Versions from space DEV and version r079

This section describes some techniques for performing analysis across data stored in multiple columns. For example, you may want to analyze combinations of height and weight. Some options:

  • Consolidate dimensions to a single metric. For example, height and weight can be combined using a BMI (body mass index) calculation. Then, use available outlier analysis capabilities in

    D s product
    rtrue
    . Below, you can review a method for bringing together similar data from multiple columns into a single column for easier analysis.

  • Flag outlier values of individual columns, perhaps giving each column a weighting factor (e.g. 0.5). Sum the outliers and their weights together.
  • Defer analysis until the data has arrived in the target system.

If you have homogeneous data across multiple columns, such as multiple individual events recorded in a single row, you can use a different method to calculate metrics. See Calculate Metrics across Columns.

In some cases, you may need to identify outliers across multiple columns of data. For example, you have a dataset containing scores from three separate tests taken by a set of individuals. Your columns may look like the following:

  • LastName
  • FirstName
  • TestScore1
  • TestScore2
  • TestScore3

You can download the Dataset-TestScores.csv dataset.Most calculations, such as standard deviation, work for a single column of data. To perform analysis across all three columns, you must reshape the above dataset to look like the following:

  • LastName
  • FirstName
  • TestNumber
  • TestScore

This steps below outline the workflow for this example. The full recipe is provided at the bottom of this section.

Steps:

  1. Load the TestScores dataset into the Transformer page. It should already be split out into five separate columns.
  2. The three columns listed side by side are data that has been organized in a pivot table. To break down this data, you must unpivot the data, which breaks down the data into a key column (containing TestScore1, TestScore2, TestScore1) and a value column, which contains individual test scores.

    D trans
    Typestep
    p01NameColumns
    p01ValueTestScore1,TestScore2,TestScore3
    p02NameGroup size
    p02Value1
    SearchTermUnpivot columns

  3. Rename the generated column of test scores to TestScore.
  4. The numeric information in the key column values can be extracted using the following:

    D trans
    p03Value`{digit}`
    Typestep
    p01NameColumn to extract from
    p01Valuekey
    p02NameOption
    p02ValueCustom text or pattern
    p03NameText to extract
    SearchTermExtract text or pattern

  5. The key2 column contains just the numeric data now. Rename this column to TestNumber. You can delete the key column now.
  6. The dataset does not contain a primary key, which field containing a unique identifier for each row. The combination of last name, first name, and test number is a unique identifier for each row in the dataset:

    D trans
    Typestep
    p01NameColumns
    p01ValueLastName,FirstName,TestNumber
    p02NameSeparator
    p02Value'-'
    SearchTermMerge columns

  7. Rename the new column to TestID. Typically, primary keys are listed as the first field in a dataset. You might want to move the column before the LastName column.
  8. You may have noticed that the data is still organized by name (first and last) and test number, so that an individual's tests are scattered throughout the dataset. To reorganize the information, you can re-aggregate the data using the following:

    D trans
    p03Value1
    Typestep
    p01NameRow labels
    p01ValueLastName,FirstName,TestNumber,TestID
    p02NameValues
    p02ValueSUM(TestScore)
    p03NameMax number of columns to create
    SearchTermPivot table

    Tip

    Tip: The above retains all instances of tests that have been taken. If you are only interested in the average test score, you can remove the TestNumber and TestID groupings and the change the SUM function to AVERAGE. In the results, you have one average for each test taker.

  9. You may want to rename the aggregation column. Your final dataset should look like the following:

D caption
typefigure
Single column of test scores

Now that your columns of data have been consolidated to a single column, you can use the single-column transforms and functions to perform analysis.

For more information on identifying outliers in this data, see Locate Outliers.