The HoeffdingNumericSplit class implements the numeric feature splitting strategy alluded to by Domingos and Hulten in the following paper: More...

Public Types
typedef NumericSplitInfo< ObservationType >	SplitInfo
	The splitting information type required by the HoeffdingNumericSplit. More...

Public Member Functions
	HoeffdingNumericSplit (const size_t numClasses, const size_t bins=10, const size_t observationsBeforeBinning=100)
	Create the HoeffdingNumericSplit class, and specify some basic parameters about how the binning should take place. More...

	HoeffdingNumericSplit (const size_t numClasses, const HoeffdingNumericSplit &other)
	Create the HoeffdingNumericSplit class, using the parameters from the given other split object. More...

size_t	Bins () const
	Return the number of bins. More...

void	EvaluateFitnessFunction (double &bestFitness, double &secondBestFitness) const
	Evaluate the fitness function given what has been calculated so far. More...

size_t	MajorityClass () const
	Return the majority class. More...

double	MajorityProbability () const
	Return the probability of the majority class. More...

size_t	NumChildren () const
	Return the number of children if this node splits on this feature. More...

template<typename Archive >
void	Serialize (Archive &ar, const unsigned int)
	Serialize the object. More...

void	Split (arma::Col< size_t > &childMajorities, SplitInfo &splitInfo) const
	Return the majority class of each child to be created, if a split on this dimension was performed. More...

void	Train (ObservationType value, const size_t label)
	Train the HoeffdingNumericSplit on the given observed value (remember that this object only cares about the information for a single feature, not an entire point). More...

Private Attributes
size_t	bins
	The number of bins. More...

arma::Col< size_t >	labels
	This holds the labels of the points before binning. More...

arma::Col< ObservationType >	observations
	Before binning, this holds the points we have seen so far. More...

size_t	observationsBeforeBinning
	The number of observations we must see before binning. More...

size_t	samplesSeen
	The number of samples we have seen so far. More...

arma::Col< ObservationType >	splitPoints
	The split points for the binning (length bins - 1). More...

arma::Mat< size_t >	sufficientStatistics
	After binning, this contains the sufficient statistics. More...

Detailed Description

template<typename FitnessFunction, typename ObservationType = double>
class mlpack::tree::HoeffdingNumericSplit< FitnessFunction, ObservationType >

The HoeffdingNumericSplit class implements the numeric feature splitting strategy alluded to by Domingos and Hulten in the following paper:

@inproceedings{domingos2000mining,
  title={{Mining High-Speed Data Streams}},
  author={Domingos, P. and Hulten, G.},
  year={2000},
  booktitle={Proceedings of the Sixth ACM SIGKDD International Conference on
      Knowledge Discovery and Data Mining (KDD '00)},
  pages={71--80}
}

The strategy alluded to is very simple: we discretize the numeric features that we see. But in this case, we don't know how many bins we have, which makes things a little difficult. This class only makes binary splits, and has a maximum number of bins. The binning strategy is simple: the split caches the minimum and maximum value of points seen so far, and when the number of points hits a predefined threshold, the cached minimum-maximum range is equally split into bins, and splitting proceeds in the same way as with the categorical splits. This is a simple and stupid strategy, so don't expect it to be the best possible thing you can do.

Template Parameters

FitnessFunction	Fitness function to use for calculating gain.
ObservationType	Type of observations in this dimension.

Definition at line 53 of file hoeffding_numeric_split.hpp.