vigra::rf::algorithms Namespace Reference

Classes
class	ClusterImportanceVisitor
class	CorrectStatus
class	Draw
class	GetClusterVariables
struct	HC_Entry
class	HClustering
class	NormalizeStatus
class	PermuteCluster
class	RFErrorCallback
class	VariableSelectionResult

Functions
template<class FeatureT , class ResponseT , class ErrorRateCallBack >
void	backward_elimination (FeatureT const &features, ResponseT const &response, VariableSelectionResult &result, ErrorRateCallBack errorcallback)
template<class FeatureT , class ResponseT >
void	cluster_permutation_importance (FeatureT const &features, ResponseT const &response, HClustering &linkage, MultiArray< 2, double > &distance)
template<class FeatureT , class ResponseT , class ErrorRateCallBack >
void	forward_selection (FeatureT const &features, ResponseT const &response, VariableSelectionResult &result, ErrorRateCallBack errorcallback)
template<class FeatureT , class ResponseT , class ErrorRateCallBack >
void	rank_selection (FeatureT const &features, ResponseT const &response, VariableSelectionResult &result, ErrorRateCallBack errorcallback)

Detailed Description

This namespace contains all algorithms developed for feature selection

Function Documentation

void vigra::rf::algorithms::forward_selection	(	FeatureT const &	features,
		ResponseT const &	response,
		VariableSelectionResult &	result,
		ErrorRateCallBack	errorcallback
	)

Perform forward selection

Parameters

features	IN: n x p matrix containing n instances with p attributes/features used in the variable selection algorithm
response	IN: n x 1 matrix containing the corresponding response
result	IN/OUT: VariableSelectionResult struct which will contain the results of the algorithm. Features between result.selected.begin() and result.pivot will be left untouched.

See Also

VariableSelectionResult

Parameters

errorcallback

IN, OPTIONAL: Functor that returns the error rate given a set of features and labels. Default is the RandomForest OOB Error.

Forward selection subsequently chooses the next feature that decreases the Error rate most.

usage:

MultiArray<2, double>     features = createSomeFeatures();
MultiArray<2, int>        labels   = createCorrespondingLabels();
VariableSelectionResult  result;
forward_selection(features, labels, result);

To use forward selection but ensure that a specific feature e.g. feature 5 is always included one would do the following

VariableSelectionResult result;
result.init(features, labels);
std::swap(result.selected[0], result.selected[5]);
result.setPivot(1);
forward_selection(features, labels, result);

See Also

VariableSelectionResult

void vigra::rf::algorithms::backward_elimination	(	FeatureT const &	features,
		ResponseT const &	response,
		VariableSelectionResult &	result,
		ErrorRateCallBack	errorcallback
	)

Perform backward elimination

Parameters

features	IN: n x p matrix containing n instances with p attributes/features used in the variable selection algorithm
response	IN: n x 1 matrix containing the corresponding response
result	IN/OUT: VariableSelectionResult struct which will contain the results of the algorithm. Features between result.pivot and result.selected.end() will be left untouched.

See Also

VariableSelectionResult

Parameters

errorcallback

IN, OPTIONAL: Functor that returns the error rate given a set of features and labels. Default is the RandomForest OOB Error.

Backward elimination subsequently eliminates features that have the least influence on the error rate

usage:

MultiArray<2, double>     features = createSomeFeatures();
MultiArray<2, int>        labels   = createCorrespondingLabels();
VariableSelectionResult  result;
backward_elimination(features, labels, result);

To use backward elimination but ensure that a specific feature e.g. feature 5 is always excluded one would do the following:

VariableSelectionResult result;
result.init(features, labels);
std::swap(result.selected[result.selected.size()-1], result.selected[5]);
result.setPivot(result.selected.size()-1);
backward_elimination(features, labels, result);

See Also

VariableSelectionResult

void vigra::rf::algorithms::rank_selection	(	FeatureT const &	features,
		ResponseT const &	response,
		VariableSelectionResult &	result,
		ErrorRateCallBack	errorcallback
	)

Perform rank selection using a predefined ranking

Parameters

features	IN: n x p matrix containing n instances with p attributes/features used in the variable selection algorithm
response	IN: n x 1 matrix containing the corresponding response
result	IN/OUT: VariableSelectionResult struct which will contain the results of the algorithm. The struct should be initialized with the predefined ranking.

See Also

VariableSelectionResult

Parameters

errorcallback

IN, OPTIONAL: Functor that returns the error rate given a set of features and labels. Default is the RandomForest OOB Error.

Often some variable importance, score measure is used to create the ordering in which variables have to be selected. This method takes such a ranking and calculates the corresponding error rates.

usage:

MultiArray<2, double>     features = createSomeFeatures();
MultiArray<2, int>        labels   = createCorrespondingLabels();
std::vector<int>        ranking  = createRanking(features);
VariableSelectionResult  result;
result.init(features, labels, ranking.begin(), ranking.end());
backward_elimination(features, labels, result);

See Also

VariableSelectionResult

void vigra::rf::algorithms::cluster_permutation_importance	(	FeatureT const &	features,
		ResponseT const &	response,
		HClustering &	linkage,
		MultiArray< 2, double > &	distance
	)

Perform hierarchical clustering of variables and assess importance of clusters

Parameters

features	IN: n x p matrix containing n instances with p attributes/features used in the variable selection algorithm
response	IN: n x 1 matrix containing the corresponding response
linkage	OUT: Hierarchical grouping of variables.
distance	OUT: distance matrix used for creating the linkage

Performs Hierarchical clustering of variables. And calculates the permutation importance measures of each of the clusters. Use the Draw functor to create human readable output The cluster-permutation importance measure corresponds to the normal permutation importance measure with all columns corresponding to a cluster permuted. The importance measure for each cluster is stored as the status() field of each clusternode

See Also

HClustering

usage:

   MultiArray<2, double>     features = createSomeFeatures();
   MultiArray<2, int>        labels   = createCorrespondingLabels();
   HClustering                linkage;
   MultiArray<2, double>    distance;
   cluster_permutation_importance(features, labels, linkage, distance)
  // create graphviz output

Draw<double, int> draw(features, labels, "linkagetree.graph");
linkage.breadth_first_traversal(draw);