Configure the Stratified Sampling component - Platform For AI - Alibaba Cloud - Platform For AI

When your training data has uneven class distributions, plain random sampling underrepresents minority classes—leading to biased models. Stratified Sampling splits a dataset into groups based on a stratification column, then draws random samples from each group independently. Use this component to improve model accuracy and training stability on imbalanced datasets.

Configure the component

Method 1: Configure on the pipeline page

Add a Stratified Sampling component to your pipeline, then set the following parameters.

Fields Setting

Parameter	Required	Description
Stratification Column	Yes	The column used to group records into strata. Each unique value defines one stratum.

Parameters Setting

Parameter	Required	Description
Sample Size	Conditional	The number of records to draw from each stratum. Must be a positive integer. At least one of Sample Size or Sampling Fraction is required. See Sampling parameters.
Sampling Fraction	Conditional	The proportion of records to draw from each stratum. Must be a floating-point number in the range (0, 1). For example, `0.2` draws 20% of records from each stratum. See Sampling parameters.
Random Seed	No	Controls the randomness of sampling. Default: `1234567`. Set a fixed seed to get reproducible results across runs. If you use the default or change the seed between runs, results may differ each time.

Tuning

Parameter	Required	Description
Cores	No	Number of CPU cores for computation. Must be a positive integer. Defaults to a system-determined value.
Memory Size per Core	No	Memory allocated to each core, in MB. Must be a positive integer in the range (1, 65536). Defaults to a system-determined value.

Sampling parameters

Sample Size and Sampling Fraction control how many records are drawn from each stratum:

Sample Size specifies an exact count per stratum. Use this when you need a fixed number of records for downstream processing—for example, set 500 to draw exactly 500 records from every stratum.
Sampling Fraction specifies a proportion. Use this when you want each stratum to contribute proportionally regardless of its size—for example, set 0.2 to draw 20% of records from every stratum.

At least one of the two must be specified. If you leave both blank, an error is returned. If you set both, Sample Size takes precedence.

Method 2: Use PAI commands

Use the SQL Script component to run PAI commands. For setup details, see SQL Script.

The following example samples 200 records from stratum A, 300 from B, and 500 from C:

PAI -name StratifiedSample
    -project algo_public
    -DinputTableName="test_input"
    -DoutputTableName="test_output"
    -DstrataColName="label"
    -DsampleSize="A:200,B:300,C:500"
    -DrandomSeed=1007
    -Dlifecycle=30;

All parameters for the StratifiedSample command:

Parameter	Required	Default	Description
`inputTableName`	Yes	—	Name of the input table.
`inputTablePartitions`	No	All partitions	Partitions to read from the input table. Supported formats: `partition_name=value` (single partition) or `name1=value1/name2=value2` (multi-level). Separate multiple partitions with commas. Example: `name1=value1,value2`.
`outputTableName`	Yes	—	Name of the output table.
`strataColName`	Yes	—	Column used as the stratification key.
`sampleSize`	No	—	Number of records to sample per stratum. If set to a positive integer, that count applies to every stratum. If set to a string, use the format `stratum0:n0,stratum1:n1` to specify counts per stratum—for example, `A:200,B:300,C:500`.
`sampleRatio`	No	—	Sampling proportion per stratum. If set to a number, must be a floating-point value between 0 and 1—for example, `0.2` draws 20% of each stratum's records. If set to a string, use the format `stratum0:r0,stratum1:r1` to specify proportions per stratum.
`randomSeed`	No	`123456`	Random seed. Must be a positive integer. Set a fixed value to make sampling reproducible across runs.
`lifecycle`	No	—	Retention period of the output table in days. Valid values: 1–3650.
`coreNum`	No	System-determined	Number of CPU cores. Must be a positive integer.
`memSizePerCore`	No	System-determined	Memory per core, in MB. Valid values: 1–65536.

At least one of sampleSize or sampleRatio must be specified. If both are set, sampleSize takes precedence.