ppforest2::Bagged< M > Struct Template Reference

Bootstrap-aggregated model wrapper. More...

#include <Bagged.hpp>

Public Types
using	Ptr = std::unique_ptr<Bagged<M>>

Public Member Functions
	Bagged (std::unique_ptr< M > model, std::vector< int > sample_indices)
bool	degenerate () const
	Whether the wrapped model reported a degenerate training run.
std::vector< int >	oob_indices (int n_total) const
	Row indices of training observations not in the bootstrap sample.
bool	operator!= (Bagged const &other) const
bool	operator== (Bagged const &other) const
	Structural equality on the wrapped model only.
types::OutcomeVector	predict (types::FeatureMatrix const &x) const
	Delegate batch prediction to the wrapped model.
types::Outcome	predict (types::FeatureVector const &x) const
	Delegate single-row prediction to the wrapped model.
types::OutcomeVector	predict_oob (types::FeatureMatrix const &x, std::vector< int > const &row_idx) const
	Predict a subset of rows (typically OOB indices).

Static Public Member Functions
template<typename Derived = M>
static Ptr	make (std::unique_ptr< Derived > model, std::vector< int > sample_indices)
	Construct a Bagged<M> and return it as a Ptr.

Public Attributes
std::unique_ptr< M >	model
	The bootstrap-trained model.
std::vector< int >	sample_indices
	Row indices (into the original training set) used to train model.

Detailed Description

template<typename M>
struct ppforest2::Bagged< M >

Bootstrap-aggregated model wrapper.

Pairs a concrete model with the row indices of the bootstrap sample it was trained on, so out-of-bag queries can recover the complementary observations. Template over M so the bagging abstraction is orthogonal to the model being aggregated — today M = Tree, but nothing in this class is tree-specific, and a future Bagged<GBTree> or Bagged<LinearModel> would reuse the exact same wrapper.

M must provide:

• predict(types::FeatureVector const&) returning a scalar outcome
• predict(types::FeatureMatrix const&) returning an OutcomeVector
• bool degenerate (field or method) for the forest retry logic
• operator==(M const&) for structural equality round-trips

Defined entirely in the header so consumers don't need an explicit instantiation list — including this file with a complete M is enough to use any Bagged<M>.

Member Typedef Documentation

Ptr

template<typename M>

using ppforest2::Bagged< M >::Ptr = std::unique_ptr<Bagged<M>>

Constructor & Destructor Documentation

Bagged()

template<typename M>

ppforest2::Bagged< M >::Bagged ( std::unique_ptr< M > model, std::vector< int > sample_indices )

inline

Member Function Documentation

degenerate()

template<typename M>

bool ppforest2::Bagged< M >::degenerate ( ) const

inline

Whether the wrapped model reported a degenerate training run.

make()

template<typename M>

template<typename Derived = M>

static Ptr ppforest2::Bagged< M >::make ( std::unique_ptr< Derived > model, std::vector< int > sample_indices )

inlinestatic

Construct a Bagged<M> and return it as a Ptr.

Sugar for the std::make_unique<Bagged<M>>(std::move(model), std::move(sample_indices)) boilerplate at every callsite that builds a bag (forest training, test fixtures, deserializers). Derived defaults to M, but lets callers wrap a subclass pointer (e.g. ClassificationTree::Ptr) without an upcast at the callsite — the cast happens here in one place.

oob_indices()

template<typename M>

std::vector< int > ppforest2::Bagged< M >::oob_indices ( int n_total ) const

inline

Row indices of training observations not in the bootstrap sample.

O(n_sample + n_total) bitmap scan: a bootstrap bag has ~63% unique entries, and vector<uint8_t> access is faster than the vector<bool> proxy or a tree-set lookup. Out-of-range indices in sample_indices are silently skipped — bootstrap sampling never produces them.

Parameters

n_total

Total number of training observations.

operator!=()

template<typename M>

bool ppforest2::Bagged< M >::operator!= ( Bagged< M > const & other ) const

inline

operator==()

template<typename M>

bool ppforest2::Bagged< M >::operator== ( Bagged< M > const & other ) const

inline

Structural equality on the wrapped model only.

sample_indices is deliberately excluded from the comparison. It records which rows were used to train this bag — bookkeeping for OOB computation, not an identity property of the model. Two bags that would produce the same predictions on every input are equal here, even if they were trained on different bootstrap samples.

Callers that need to assert sample-indices round-trip must compare sample_indices directly (see Json.test.cpp's round-trip test).

predict() [1/2]

template<typename M>

types::OutcomeVector ppforest2::Bagged< M >::predict ( types::FeatureMatrix const & x ) const

inline

Delegate batch prediction to the wrapped model.

predict() [2/2]

template<typename M>

types::Outcome ppforest2::Bagged< M >::predict ( types::FeatureVector const & x ) const

inline

Delegate single-row prediction to the wrapped model.

predict_oob()

template<typename M>

types::OutcomeVector ppforest2::Bagged< M >::predict_oob ( types::FeatureMatrix const & x, std::vector< int > const & row_idx ) const

inline

Predict a subset of rows (typically OOB indices).

The returned vector has the same size as row_idx; element i is the wrapped model's prediction for row row_idx[i] of x.

Member Data Documentation

model

template<typename M>

std::unique_ptr<M> ppforest2::Bagged< M >::model

The bootstrap-trained model.

sample_indices

template<typename M>

std::vector<int> ppforest2::Bagged< M >::sample_indices

Row indices (into the original training set) used to train model.

---

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

• core/src/models/Bagged.hpp