magpie.attributes.selectors

Class RandomForestAttributeSelector

java.lang.Object

magpie.attributes.selectors.BaseAttributeSelector

magpie.attributes.selectors.PythonBasedAttributeSelector

magpie.attributes.selectors.RandomForestAttributeSelector

All Implemented Interfaces:

java.io.Serializable, java.lang.Cloneable, Commandable, Options, Printable

public class RandomForestAttributeSelector
extends PythonBasedAttributeSelector

Select attributes using the feature-importance scores from RandomForest.

This selector works by iteratively selecting only the best-performing attributes as selected by a RandomForest model, and then repeating the process on the reduced set of attributes.

A technical note you should be aware of: This selection makes no assurances that the selected attributes are not correlated with each other. You may want to follow up the selection of a subset with a tool that does ensure selected attributes are not correlated.

Procedure:

1. Determine a constant factor, f, such that f^n times the original number of attributes equals the target number of attributes, where n is the number of loop steps
2. Train a RandomForest model on the data
3. Identify the top f% of attributes based on the "feature importance" score from the RandomForest model
4. Eliminate all other attributes from the dataset
5. If the number of remaining attributes is greater than the target number, repeat from Step 2

Usage: -num_attr <attrs> [-num_steps <steps>] [-num_trees <trees>]
attrs: Number of attributes to select
steps: Number of selection steps to employ (default = 8)
trees: Number of trees to use in RandomForest (default = 100)

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
protected int	NumAttributes Number of attributes to select.
protected int	NumSteps Number of selection steps
protected int	NumTrees Number of trees to use in RandomForest

Fields inherited from class magpie.attributes.selectors.PythonBasedAttributeSelector

Debug

Fields inherited from class magpie.attributes.selectors.BaseAttributeSelector

trained

Constructor Summary

Constructors

Constructor and Description
RandomForestAttributeSelector()

Method Summary

Modifier and Type	Method and Description
protected java.util.List<java.lang.String>	assembleSystemCall(java.io.File codePath, Dataset data) Prepare the system call with all command-line arguments
protected java.lang.String	getScriptPath() Get the path to the Python script to be run.
java.lang.String	printDescription(boolean htmlFormat) Print full name of object, and a simple description of the options.
java.lang.String	printUsage() Print out required format for options.
void	setNumAttributes(int numAttributes) Set the target number of attributes to be selected
void	setNumSteps(int numSteps) Number of iterations during attribute selection
void	setNumTrees(int numTrees) Number of trees for the RandomForest model
void	setOptions(java.util.List<java.lang.Object> Options) Set any options for this object.

Methods inherited from class magpie.attributes.selectors.PythonBasedAttributeSelector

spawnStderrReader, train_protected

Methods inherited from class magpie.attributes.selectors.BaseAttributeSelector

about, applyAttributeSelection, clone, getSelectionNames, getSelections, isTrained, printCommand, printSelections, run, runCommand, train

Methods inherited from class java.lang.Object

equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

NumAttributes

protected int NumAttributes

Number of attributes to select.

NumSteps

protected int NumSteps

Number of selection steps

NumTrees

protected int NumTrees

Number of trees to use in RandomForest

Constructor Detail

RandomForestAttributeSelector

public RandomForestAttributeSelector()

Method Detail

setOptions

public void setOptions(java.util.List<java.lang.Object> Options)
                throws java.lang.Exception

Description copied from interface: Options

Set any options for this object. If parsing the options fails, return an Exception with the usage as its message.

Parameters:

Options - Array of options as Objects - can be null

Throws:

java.lang.Exception - if problem with inputs

printUsage

public java.lang.String printUsage()

Description copied from interface: Options

Print out required format for options. For consistency, should start with "Usage: ".

Returns:

Usage for this function

getScriptPath

protected java.lang.String getScriptPath()

Description copied from class: PythonBasedAttributeSelector

Get the path to the Python script to be run.

Specified by:

getScriptPath in class PythonBasedAttributeSelector

Returns:

Path

setNumAttributes

public void setNumAttributes(int numAttributes)

Set the target number of attributes to be selected

Parameters:

numAttributes - Target number of attributes

setNumSteps

public void setNumSteps(int numSteps)

Number of iterations during attribute selection

Parameters:

numSteps - Desired number of steps

setNumTrees

public void setNumTrees(int numTrees)

Number of trees for the RandomForest model

Parameters:

numTrees - Desired number of trees

assembleSystemCall

protected java.util.List<java.lang.String> assembleSystemCall(java.io.File codePath,
                                                              Dataset data)

Description copied from class: PythonBasedAttributeSelector

Prepare the system call with all command-line arguments

Specified by:

assembleSystemCall in class PythonBasedAttributeSelector

Parameters:

codePath - Path to executable or script to be run

data - Dataset being used to train attribute selector

Returns:

Command to be executed

printDescription

public java.lang.String printDescription(boolean htmlFormat)

Description copied from interface: Printable

Print full name of object, and a simple description of the options. This description should print enough information to reproduce how this object was created, but not necessarily enough to run it (e.g., coefficients for the model).

Example: For a model training a separate WekaRegression for intermetallics

magpie.models.regression.SplitRegression

Splitter: AllMetalsSplitter
All Metals: magpie.models.regression.WekaRegression trees.REPTree
Contains Nonmetal: magpie.regression.LASSORegreession -maxterms 2

Specified by:

printDescription in interface Printable

Overrides:

printDescription in class BaseAttributeSelector

Parameters:

htmlFormat - Whether format for output to an HTML page (e.g., <div> to create indentation) or for printing to screen.

Returns:

String describing the model

See Also:

#printModel()