Configure PS-SMART Multiclass Classification - Platform For AI

PS-SMART trains multiclass classification models using a GBDT-based algorithm that scales to billions of samples and thousands of nodes.

Limitations

Target column accepts only numeric types. Convert STRING data in MaxCompute tables before training. Example: Convert Good/Medium/Bad to 0/1/2.
For KV format data, feature IDs must be positive integers and feature values must be real numbers. Convert string feature IDs using the serialize component. Apply feature engineering (such as discretization) to categorical string values.
Component handles hundreds of thousands of features but consumes significant resources and runs slowly. GBDT algorithms perform best with continuous features. Apply one-hot encoding to categorical features to filter low-frequency ones. Avoid discretizing continuous features.
PS-SMART introduces randomness through data sampling (data_sample_ratio), feature sampling (fea_sample_ratio), histogram approximation, and sketch merging order. Tree structures may differ across distributed workers, but model performance remains similar. Identical data and parameters may produce inconsistent results across runs.
Increasing computing cores accelerates training. Training starts only after all servers acquire necessary resources. Requesting more resources during cluster congestion increases wait time.

Configure the component

Method 1: Use GUI

Add PS-SMART multiclass classification to Designer workflow and configure parameters:

Parameter type	Parameter	Description
Fields setting	Is sparse format	Sparse format uses spaces to separate KV pairs and colons to separate key and value. Example: 1:0.3 3:0.9.
	Feature columns	Feature columns from input table. For dense format, select only numeric columns (BIGINT or DOUBLE). For sparse KV format where both key and value are numeric, select only STRING columns.
	Label column	Label column from input table. Accepts STRING and numeric types. Internal storage requires numeric types, such as 0 and 1 for binary classification.
	Weight column	Weight column for each sample row. Accepts only numeric types.
Parameters setting	Number of classes	Number of classes for multiclass classification. If set to n, label column values must range from 0 to n-1.
	Evaluation metric type	Evaluation metric types: multiclass negative log likelihood and multiclass classification error.
	Number of trees	Number of trees (positive integer). Training time increases proportionally.
	Maximum tree depth	Maximum tree depth. Default: 5 (allows up to 32 leaf nodes).
	Data sampling ratio	Proportion of data sampled when building each tree. Creates weak learners and accelerates training.
	Feature sampling ratio	Proportion of features sampled when building each tree. Creates weak learners and accelerates training.
	L1 penalty coefficient	L1 penalty coefficient. Controls leaf node size distribution. Larger values produce more uniform distribution. Increase to prevent overfitting.
	L2 penalty coefficient	L2 penalty coefficient. Controls leaf node size distribution. Larger values produce more uniform distribution. Increase to prevent overfitting.
	Learning rate	Learning rate. Range: (0,1).
	Approximate sketch precision	Quantile threshold for sketch construction. Smaller values create more buckets. Default: 0.03. Manual configuration not required.
	Minimum split loss change	Minimum loss reduction required to split a node. Larger values produce more conservative splits.
	Number of features	Number of features or maximum feature ID. Used for resource estimation.
	Global bias	Initial prediction value for all samples. Also known as base score.
	Random number generator seed	Random seed (integer).
	Feature importance type	Feature importance calculation types: Split count: Number of times feature is used for splitting. Information gain: Information gained from feature. Sample coverage: Number of samples covered by feature at split nodes.
Execution tuning	Number of cores	Number of computing cores. System allocates automatically by default.
Execution tuning	Memory size per core	Memory per core (MB). System allocates automatically. Manual configuration rarely needed.

Method 2: Use PAI commands

Call PAI commands from a SQL script component. For details, see Scenario 4: Execute PAI commands in a SQL script component.

--Train
PAI -name ps_smart
    -project algo_public
    -DinputTableName="smart_multiclass_input"
    -DmodelName="xlab_m_pai_ps_smart_bi_545859_v0"
    -DoutputTableName="pai_temp_24515_545859_2"
    -DoutputImportanceTableName="pai_temp_24515_545859_3"
    -DlabelColName="label"
    -DfeatureColNames="features"
    -DenableSparse="true"
    -Dobjective="multi:softprob"
    -Dmetric="mlogloss"
    -DfeatureImportanceType="gain"
    -DtreeCount="5"
    -DmaxDepth="5"
    -Dshrinkage="0.3"
    -Dl2="1.0"
    -Dl1="0"
    -Dlifecycle="3"
    -DsketchEps="0.03"
    -DsampleRatio="1.0"
    -DfeatureRatio="1.0"
    -DbaseScore="0.5"
    -DminSplitLoss="0"
--Predict
PAI -name prediction
    -project algo_public
    -DinputTableName="smart_multiclass_input";
    -DmodelName="xlab_m_pai_ps_smart_bi_545859_v0"
    -DoutputTableName="pai_temp_24515_545860_1"
    -DfeatureColNames="features"
    -DappendColNames="label,features"
    -DenableSparse="true"
    -DkvDelimiter=":"
    -Dlifecycle="28"

Module	Parameter	Required	Default	Description
Data parameters	featureColNames	Yes	None	Feature columns from input table. For dense format, select only numeric columns (BIGINT or DOUBLE). For sparse KV format where key and value are numeric, select only STRING columns.
	labelColName	Yes	None	Label column from input table. Accepts STRING and numeric types. Internal storage requires numeric types. For multiclass classification, values must range from 0 to n-1, where n is the number of classes.
	weightCol	No	None	Weight column for each sample row. Accepts only numeric types.
	enableSparse	No	false	Whether data is in sparse format. Values: true or false. Sparse format uses spaces to separate KV pairs and colons to separate key and value. Example: 1:0.3 3:0.9.
	inputTableName	Yes	None	Name of input table.
	modelName	Yes	None	Name of output model.
	outputImportanceTableName	No	None	Name of output table for feature importance.
	inputTablePartitions	No	None	Format: ds=1/pt=1.
	outputTableName	No	None	Output table in MaxCompute. Binary format, not directly readable. Accessible only through SMART prediction component.
	lifecycle	No	3	Lifecycle of output table (in days).
Algorithm parameters	classNum	Yes	None	Number of classes for multiclass classification. If set to n, label column values must range from 0 to n-1.
	objective	Yes	None	Objective function type. For multiclass classification training, use multi:softprob.
	metric	No	None	Evaluation metric type for training dataset. Output written to stdout in Logview Coordinator area. Types: mlogloss: Multiclass negative log likelihood (corresponds to GUI option). merror: Multiclass classification error (corresponds to GUI option).
	treeCount	No	1	Number of trees (positive integer). Training time increases proportionally.
	maxDepth	No	5	Maximum tree depth. Range: 1 to 20.
	sampleRatio	No	1.0	Data sampling ratio. Range: (0,1]. Value 1.0 disables sampling.
	featureRatio	No	1.0	Feature sampling ratio. Range: (0,1]. Value 1.0 disables sampling.
	l1	No	0	L1 penalty coefficient. Larger values produce more uniform leaf node distribution. Increase to prevent overfitting.
	l2	No	1.0	L2 penalty coefficient. Larger values produce more uniform leaf node distribution. Increase to prevent overfitting.
	shrinkage	No	0.3	Learning rate. Range: (0,1).
	sketchEps	No	0.03	Quantile threshold for sketch construction. Bucket count: O(1.0/sketchEps). Smaller values create more buckets. Use default value. Manual configuration not required. Range: (0,1).
	minSplitLoss	No	0	Minimum loss reduction required to split a node. Larger values produce more conservative splits.
	featureNum	No	None	Number of features or maximum feature ID. Used for resource estimation.
	baseScore	No	0.5	Initial prediction value for all samples. Also known as global bias.
	randSeed	No	None	Random seed (integer).
	featureImportanceType	No	gain	Feature importance calculation types: weight: Split count - number of times feature is used for splitting. gain: Information gain from feature. cover: Sample coverage - number of samples covered by feature at split nodes.
Tuning parameters	coreNum	No	System allocated	Number of computing cores. Larger values accelerate execution.
Tuning parameters	memSizePerCore	No	System allocated	Memory per core in MB.

Model deployment

To deploy model generated by PS-SMART component as online service, add General-purpose Model Export component downstream of PS-SMART component. Configure component parameters same way as other PS-series components. For more information, see General-purpose Model Export.

Upon successful execution, deploy model service on PAI-EAS Model Online Service page. For more information, see Deploy a service in the console.