FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR > Class Template Reference

Multinomial ("bag of words") model: per-class term-count sums and the Laplace-smoothed term probabilities P(term|class) derived from them. More...

#include <model_multinom.hpp>

Inheritance diagram for FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >:

Public Types
typedef std::shared_ptr< DATAACCESSOR >	PDataAccessor
	shared pointer to the data accessor supplying train blocks
typedef std::shared_ptr< SUBSET >	PSubset
	shared pointer to a feature subset, as consumed by learn()/narrow_to()
Public Types inherited from FST::Model< SUBSET, DATAACCESSOR >
typedef std::shared_ptr< DATAACCESSOR >	PDataAccessor
	shared pointer to the data accessor supplying training data
typedef std::shared_ptr< SUBSET >	PSubset
	shared pointer to a feature subset, as consumed by narrow_to()

Public Member Functions
	Model_Multinomial (const Model_Multinomial &mm)
	Deep-copies all learned buffers (sufficient statistics, theta/MI/IB caches, index maps).
virtual void	learn (PDataAccessor da) override
	Fits the model's parameters from da's training data, over all its features. Concrete models may additionally offer a (da, subset) convenience overload to learn at a lower dimensionality directly - that is not part of this interface, so generic code must not assume it exists. narrow_to()/denarrow() do NOT relearn – they only change which features of an ALREADY-learned model are exposed to subsequent evaluation.
virtual void	learn (PDataAccessor da, const PSubset sub)
	Learns the model restricted to sub's selected features. Computes only the sufficient statistics (_Nsuminclass, class priors _Pc, _Pc_d) via compute_Nsuminclass(); it does NOT fill _theta/_MI/_IB – call compute_theta() (and, for the full feature set, compute_MI()/compute_IB()) explicitly afterwards. Reallocates buffers when the class count or sub's raw dimensionality exceeds current capacity; otherwise reuses them. Leaves the model unnarrowed (_d=_n).
void	narrow_to (const PSubset sub) override
void	denarrow () override
	Undoes narrow_to(): model access reverts to the full dimensionality it was learned at.
void	compute_theta ()
	Computes the Laplace-smoothed term-conditional probabilities _theta = P(feature|class) and _doc_avg_length, over the currently exposed dimensionality (_d, via narrow_to()/denarrow()).
void	compute_MI ()
	Computes per-feature Mutual Information (fills _MI, full _n-length).
void	compute_IB ()
	Computes per-feature "Individual Bhattacharyya" (fills _IB, full _n-length): the class-pair-averaged Bhattacharyya-like coefficient of each single feature's term distribution, weighted by _Pc_d and scaled by _doc_avg_length.
DIMTYPE	get_n_max () const
	allocated buffer capacity (dimensionality upper bound)
DIMTYPE	get_n () const
	"full feature set" dimensionality the model was learned at
DIMTYPE	get_d () const
	dimensionality currently exposed (post narrow_to()/denarrow())
DIMTYPE	get_classes () const
	number of classes, as set by the last learn() call
REALTYPE	get_doc_avg_length () const
	average total term length per training pattern; valid after compute_theta()
REALTYPE	get_Pc (const DIMTYPE c) const
	class prior P(class=c), estimated in learn() from class sizes

Protected Member Functions
const std::unique_ptr< REALTYPE[]> &	get_theta () const
	raw P(term|class) buffer, class-major, valid at the dimensionality last passed to compute_theta()
const std::unique_ptr< REALTYPE[]> &	get_IB () const
	raw per-feature Individual Bhattacharyya buffer (full _n length), valid after compute_IB()
const std::unique_ptr< REALTYPE[]> &	get_MI () const
	raw per-feature Mutual Information buffer (full _n length), valid after compute_MI()
void	compute_Nsuminclass (PDataAccessor da)
	Sums raw per-class, per-(learn-subspace)-feature term counts into _Nsuminclass and sets _allpatterns. Called from learn(da,sub); the subspace is the one recorded in _learn_index.

Protected Attributes
DIMTYPE	_classes
	number of classes; initialized in learn(). NOTE: expressions like _classes*_n*_n are evaluated in std::size_t where they occur, so they do not overflow DIMTYPE
DIMTYPE	_n_max
	allocated buffer capacity (dimensionality upper bound); see the NOTE above
DIMTYPE	_n
std::unique_ptr< DATATYPE[]>	_Nsuminclass
	per-class, per-learn-subspace-feature raw term-count sums (class-major, stride _n); filled by compute_Nsuminclass()
std::unique_ptr< REALTYPE[]>	_theta
	P(term|class), Laplace-smoothed; class-major, stride = the _d in effect at the last compute_theta() call.
std::unique_ptr< REALTYPE[]>	_Pc
	per-class prior P(class), estimated in learn() from da->getClassSize()/getClassSizeSum()
std::unique_ptr< REALTYPE[]>	_Pc_d
	per-class share of total learn-subspace term length; computed in learn(), consumed by compute_MI()/compute_IB()
std::unique_ptr< REALTYPE[]>	_Pv
	per-feature marginal term probability scratch, used internally by compute_MI()
std::unique_ptr< REALTYPE[]>	_MI
	per-feature (full _n-length) Mutual Information scores, filled by compute_MI()
std::unique_ptr< REALTYPE[]>	_IB
	per-feature (full _n-length) Individual Bhattacharyya scores, filled by compute_IB()
DIMTYPE	_d
	"virtual subset" dimensionality currently exposed (<=_n), via narrow_to()/denarrow()
std::unique_ptr< DIMTYPE[]>	_index
	maps feature subset indexes to raw (full set) feature indexes, used for narrow()ing
IDXTYPE	_allpatterns
	total training pattern count summed over all classes; set by compute_Nsuminclass() (called from learn())
REALTYPE	_doc_avg_length
	average total term length per training pattern; computed in compute_theta(), needed by compute_IB()
std::unique_ptr< DIMTYPE[]>	_learn_index
	maps the to-be-learned model's subspace positions to raw (full-set) feature indices; set by learn(da,sub)

Friends
template<class, class, class, class, class, class, class>
class	Criterion_Multinomial_Bhattacharyya
template<class, class, class, class, class, class, class>
class	Criterion_Multinomial_MI
template<class, class, class, class, class, class, class>
class	Classifier_Multinomial_NaiveBayes

Detailed Description

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
class FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >

Multinomial ("bag of words") model: per-class term-count sums and the Laplace-smoothed term probabilities P(term|class) derived from them.

For every class the per-feature term counts of all training patterns are summed (compute_Nsuminclass()); compute_theta() turns those sums into the smoothed conditional probabilities

theta(f|c) = (1 + N(f,c)) / (d + sum over the exposed f' of N(f',c))

that is, add-one (Laplace) smoothing over the ACTIVE vocabulary: the d features currently exposed by narrow_to(), not the full feature set, so the probabilities of a feature subset are normalized within that subset. Class priors P(c) are estimated from the class sizes of the WHOLE data set the accessor holds (not the train part alone; identical under stratified splitting), while the term counts come from the TRAIN part. compute_MI() and compute_IB() add the per-feature Mutual Information and "Individual Bhattacharyya" rankings consumed by the multinomial criteria.

Unlike Model_Normal this model carries no mean/covariance and makes no continuity assumption – it is the discrete counterpart, intended for count data such as document term frequencies.

Note

The data values are assumed to be non-negative term counts. Nothing verifies it; negative or fractional values make the sums and the smoothed probabilities meaningless.

Call order is learn() (or learn(da,sub)), then compute_theta(), then optionally compute_MI() / compute_IB() – and those two require compute_theta() to have been run at the FULL (denarrow()ed) feature set. The order is not enforced; calling out of order silently uses stale buffers.

Member Function Documentation

compute_IB()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::compute_IB

(

)

Computes per-feature "Individual Bhattacharyya" (fills _IB, full _n-length): the class-pair-averaged Bhattacharyya-like coefficient of each single feature's term distribution, weighted by _Pc_d and scaled by _doc_avg_length.

Note

Same full-feature-set precondition on _theta as compute_MI().

compute_MI()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::compute_MI

(

)

Computes per-feature Mutual Information (fills _MI, full _n-length).

Note

Requires compute_theta() to have been called at the FULL (denarrow()ed) feature set – this reads _theta densely over 0.._n per class, not through _index.

compute_theta()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::compute_theta

(

)

Computes the Laplace-smoothed term-conditional probabilities _theta = P(feature|class) and _doc_avg_length, over the currently exposed dimensionality (_d, via narrow_to()/denarrow()).

Note

Requires learn()/learn(da,sub) to have populated _Nsuminclass/_allpatterns first. compute_MI()/compute_IB() additionally require this to have been called while denarrow()ed (_d==_n) – they index _theta over the full _n range directly.

denarrow()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::denarrow ( )

overridevirtual

Undoes narrow_to(): model access reverts to the full dimensionality it was learned at.

Implements FST::Model< SUBSET, DATAACCESSOR >.

learn()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::learn ( PDataAccessor da )

overridevirtual

Fits the model's parameters from da's training data, over all its features. Concrete models may additionally offer a (da, subset) convenience overload to learn at a lower dimensionality directly - that is not part of this interface, so generic code must not assume it exists. narrow_to()/denarrow() do NOT relearn – they only change which features of an ALREADY-learned model are exposed to subsequent evaluation.

Implements FST::Model< SUBSET, DATAACCESSOR >.

narrow_to()

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

void FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::narrow_to ( const PSubset sub )

overridevirtual

Restricts subsequent model access to sub's selected features, without relearning (analogous to Distance::narrow_to() – see there for the _index[] mechanism concrete models typically reuse).

Implements FST::Model< SUBSET, DATAACCESSOR >.

Member Data Documentation

_n

template<typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>

DIMTYPE FST::Model_Multinomial< DATATYPE, REALTYPE, IDXTYPE, DIMTYPE, SUBSET, DATAACCESSOR >::_n

protected

"full feature set" dimensionality the model was learned at (<=_n_max); initialized in learn(). NOTE: expressions like _classes*_n*_n are evaluated in std::size_t where they occur, so they do not overflow DIMTYPE

---

The documentation for this class was generated from the following file:

• /Users/psms/dev/devtest1/_src_criteria/model_multinom.hpp