Skip to main content

Seemingly Unrelated Regression

Synopsis​

This operator performs regressions on coherent ExampleSets with different labels and will take into account the correlation of residuals to build the model. It creates a composite regression model optimized for predicting all label attributes at the same time.

Description​

This operator creates a model that consists of multiple regressions targeting different label attributes for the same set of examples. Once each single linear regression is built for each input ExampleSet (delivered at unrelated example sets ports), the covariance of the residuals is used to improve the quality of predictions that are influenced by effects not captured by the attributes.

The first port is dedicated for the original ExampleSet, which must contain the union of all attributes delivered at unrelated example sets input ports. Moreover, all input ExampleSets must have the same number of examples.

Practically, all seemingly unrelated input ExampleSets should have the same examples as the original one, but only contain a subset of its attributes and its own dedicated label.

Please note that whenever a Seemingly Unrelated Regression model is applied, the prediction attributes will not have prediction roles, as only one attribute can have this role at a time. You can add prediction role to a single attribute by using Set Role operator.

Input​

training set​

unrelated example sets​

Output​

model​

example set​

This is an example set output port

Parameters​

feature selection​

The feature selection method used during regression.

alpha​

This is the significance level of the t-test.

max iterations​

The maximal number of rounds for iterative selection.

forward alpha​

This is the alpha level for the used t-test for selecting attributes.

backward alpha​

This is the alpha level for the used t-test for deselecting attributes.

eliminate colinear features​

Indicates if the algorithm should try to delete colinear features during the regression.

min tolerance​

The minimum tolerance for the removal of colinear features.

use bias​

Indicates if an intercept value should be calculated.

ridge​

The ridge parameter used for ridge regression. A value of zero switches to ordinary least squares estimate.