Discretize ML Features with Binning in PAI - Platform for AI - Alibaba Cloud - Platform For AI

The Binning component performs feature discretization by converting continuous data into multiple discrete intervals. This component supports equal frequency binning, equal width binning, and automated binning.

Configure the component

You can configure the Binning component using one of the following methods.

Method 1: Use the UI

Configure the component parameters on the Designer workflow page.

Tab	Parameter	Description
Fields Setting	Feature Columns	Supports STRING, BIGINT, and DOUBLE data types.
	Label Column	Only binary classification is supported.
	Positive Value	This parameter applies only when a Label Column is specified.
	Binning Parameter Source	Valid values: Parameters in Parameter Settings and Manual Binning or Custom JSON.
	Reserve Unselected Feature Columns	If you use custom binning and select Yes, columns not selected in Feature Columns are retained. Otherwise, they are dropped.
	Upload Binning and Constraint JSON	This parameter applies only when Binning Parameter Source is set to Manual Binning or Custom JSON.
Parameters Setting	Bins	If you set this parameter to 10, continuous features are discretized into 10 intervals.
	Custom Bins	You can specify the number of bins for one or more columns, which overrides the global number of bins. If you customize the number of bins for a column that is not in your column selection, that column is also included in the calculation. For example, if you select columns col0 and col1, define custom binning as col0:3,col2:5, and the global Number of Bins is 10, the calculation is performed by using the configuration col0:3,col1:10,col2:5. Specify the value in the format: ColumnName1:NumberOfBins,ColumnName2:NumberOfBins.
	Custom Discrete Value Count Threshold	Specifies a custom frequency threshold for discrete values in one or more columns. Values with a frequency below this threshold are assigned to the else bin. This setting overrides the global Discrete Value Count Threshold. Use the format ColumnName:Threshold, for example, col0:3.
	Interval Type	Supports left-open, right-closed or left-closed, right-open intervals.
	Binning Mode	Valid values: Equal Frequency, Equal Width, and Automatic Binning.
	Discrete Value Count Threshold	If the frequency of a discrete value is less than this threshold, the value is assigned to the else bin.
Tuning	Cores	The number of CPU cores for the job. By default, the system allocates cores automatically.
Tuning	Memory Size per Core	The memory per core, in MB. By default, the system allocates memory automatically.

Method 2: Use a PAI command

Use a PAI command to configure the component parameters. You can run PAI commands by using the SQL Script component. For more information, see SQL Script.

PAI -name binning
    -project algo_public
    -DinputTableName=input
    -DoutputTableName=output

Parameter	Description	Required	Default
inputTableName	The name of the input table.	Yes	N/A
outputTableName	The name of the output table.	Yes	N/A
selectedColNames	The columns from the input table to bin.	No	All columns except the label column. If a label column is not specified, all columns are selected.
labelColName	The label column.	No	N/A
validTableName	The name of the validation table for the `auto` binning mode. This parameter is required when `binningMethod` is `auto`.	No	Empty
validTablePartitions	The partitions to select from the validation table.	No	The full table.
inputTablePartitions	The partitions to select from the input table.	No	The full table.
inputBinTableName	The input binning table.	No	N/A
selectedBinColNames	The columns to select from the binning table.	No	Empty
positiveLabel	The class value for positive samples in the output.	No	1
nDivide	The number of bins. The value must be a positive integer.	No	10
colsNDivide	The number of bins for custom columns, such as col0:3,col2:5. If a column specified in colsNDivide is not in selectedColNames, the additional column is also included in the calculation. For example, if selectedColNames is col0,col1, colsNDivide is col0:3,col2:5, and nDivide is 10, the calculation is performed based on col0:3,col1:10,col2:5.	No	Empty
isLeftOpen	Specifies the interval type. Valid values: {true}: Left-open, right-closed. {false}: Left-inclusive and right-exclusive.	No	true
stringThreshold	The frequency threshold for discrete values to be assigned to the else bin.	No	N/A
colsStringThreshold	The threshold for custom columns is the same as colsNDivide.	No	Empty
binningMethod	The binning mode. Valid values: quantile: Equal frequency binning. bucket: equal width binning. auto: In quantile mode, automatically selects monotonic binning.	No	quantile
lifecycle	The lifecycle of the output table in days. The value must be a positive integer.	No	N/A
coreNum	The number of cores. The value must be a positive integer.	No	Automatically calculated by the system.
memSizePerCore	The memory per core in MB. The value must be a positive integer.	No	Automatically calculated by the system.

The binning constraint feature is designed for use with the Scorecard Training component. In this process, binning serves as a feature engineering step to discretize features into dummy variables. You can then constrain the weight of each dummy variable as follows:

Ascending order: Constrains the weights of the dummy variables for a feature to increase as their index values increase.
Descending order: Constrains the weights of the dummy variables for a feature to decrease as their index values increase.
Same weight: Forces selected dummy variables for a feature to have the same weight.
Zero weight: Constrains the weight of a dummy variable to be zero.
Specific weight: Constrains the weight of a dummy variable to a specific floating-point value.
WOE order: Constrains the weights of the dummy variables for a feature to increase as their Weight of Evidence (WOE) values increase.

View results

Once a workflow containing the Binning component has run, right-click the Binning component on the canvas and select Binning from the shortcut menu.
On the variable list page, you can view the Bins, Type, and IV for each variable.
Click the name of a variable, such as f1, to open the f1-binning details page.

On this page, you can Merge or Split bins and add constraints to them.

Note
Constraints apply only to a subsequent Scorecard Training component. You can ignore them if you are not using one.

The page supports two views: List and Chart. The list view displays the Index, Label (the bin interval, such as (-inf,4.8]), Constrain (the constraint operator and value), WOE, Number (total, positive, and negative counts), and Rate for each bin. The Custom Binning panel on the right allows you to specify bins by data type, set boundary ranges, and configure constraint types, including Ascending order (< bin), Descending order (> bin), Same weight (= bin), Zero weight (= 0), Specific weight (= weight), and WOE order.