Principal component analysis (PCA)-Platform For AI(PAI)-阿里云帮助中心

Principal Component Analysis (PCA) is a multivariate statistical method for exploring the internal structure of multiple variables and examining their correlations. PCA performs dimensionality reduction by deriving a few uncorrelated principal components from the original variables. These components retain as much information as possible from the original data and serve as new composite indicators.

Limitations

The Principal Component Analysis (PCA) component performs dimensionality reduction and noise reduction. It supports only dense data.

Configuration

Method 1: Visual interface

In Machine Learning Designer, add the Principal Component Analysis (PCA) component to your pipeline and configure its parameters in the pane on the right.

Tab	Parameter	Description
Fields Setting	Feature Columns	The columns from the input table to use for analysis.
Fields Setting	Appended Columns	The columns to append to the dimensionally-reduced output table.
Parameters Setting	Information Retention Ratio	The proportion of information from the original data to retain after dimensionality reduction.
	Feature Decomposition Mode	The method used for feature decomposition. Valid values: CORR COVAR_SAMP COVAR_POP
	Data Conversion Method	The method used to transform data into principal components. Valid values: Simple Sub-Mean Normalization
Tuning	Lifecycle	The lifecycle of the output table. Must be a positive integer.
	Number of Cores	Used with the Memory Size per Core (MB) parameter. The value must be a positive integer in the range [1, 9999].
	Memory Size per Core (MB)	Unit: MB. The value must be a positive integer in the range [1024, 65536].

Method 2: PAI commands

Alternatively, run PAI (Machine Learning Platform for AI) commands in a SQL Script component to configure the Principal Component Analysis (PCA) component. For more information, see SQL Script.

PAI -name PrinCompAnalysis
    -project algo_public
    -DinputTableName=bank_data
    -DeigOutputTableName=pai_temp_2032_17900_2
    -DprincompOutputTableName=pai_temp_2032_17900_1
    -DselectedColNames=pdays,previous,emp_var_rate,cons_price_idx,cons_conf_idx,euribor3m,nr_employed
    -DtransType=Simple
    -DcalcuType=CORR
    -DcontriRate=0.9;

Parameter	Required	Default	Description
inputTableName	Yes	None	The input table for principal component analysis.
selectedColNames	Yes	None	The names of the columns in the input table to be used for analysis. Separate multiple column names with a comma. Only columns of the INT and DOUBLE data types are supported.
eigOutputTableName	Yes	None	The output table for the feature vectors and eigenvalues.
princompOutputTableName	Yes	None	The output table containing the results after dimensionality reduction and noise reduction.
transType	No	Simple	The method used to transform data into principal components. Valid values: Simple Sub-Mean Normalization
calcuType	No	CORR	The method used for feature decomposition. Valid values: CORR COVAR_SAMP COVAR_POP
contriRate	No	0.9	The proportion of information to retain after dimensionality reduction. The value must be in the range (0, 1).
remainColumns	No	None	The columns from the original table to include in the dimensionally-reduced output table.
coreNum	No	System-assigned	The number of cores to use. This parameter works in conjunction with the memSizePerCore parameter. The value must be a positive integer in the range [1, 9999].
memSizePerCore	No	System-assigned	The memory size for each core, in MB. The value must be a positive integer in the range [1024, 65536].
lifecycle	No	None	The lifecycle of the output table. Must be a positive integer.

Examples

Example outputs

This output table contains four principal component columns: prin0, prin1, prin2, and prin3. Each row represents the principal component values for a sample after dimensionality reduction.
Eigenvalue and feature vector table: This table contains four principal components (prin0 to prin3). The columns show the feature vector components (pdays, previous, emp_var_rate, cons_price_idx, cons_conf_idx, euribor3m, and nr_employed), and the corresponding eigenvalue, contributionrate (contribution rate), and sumcontributionrate (cumulative contribution rate).