FST::Metric_Spec< RETURNTYPE > Struct Template Reference

configuration of the wrapper metric; immutable during search, copied by criterion clones (thread-safe, RNG-free) More...

#include <wrapper_metrics.hpp>

Public Member Functions
bool	is_default () const
	true iff this spec is the plain unweighted accuracy path (f1_positive is irrelevant unless metric==f1, so it stays out of this test)
bool	is_regression_metric () const
	true for the three regression metrics (r2/neg_rmse/neg_mae), false for the classification metrics

Public Attributes
Wrapper_Metric	metric = Wrapper_Metric::accuracy
	which metric to compute; accuracy = the default path served by Classifier::test()
std::vector< RETURNTYPE >	class_weights
unsigned int	f1_positive = 1
	class index the f1 metric scores as positive (two-class data; --f1-positive). Affects f1 ONLY - f1-macro averages over all classes, and auc keeps class 1 positive by its own convention

Detailed Description

template<typename RETURNTYPE>
struct FST::Metric_Spec< RETURNTYPE >

configuration of the wrapper metric; immutable during search, copied by criterion clones (thread-safe, RNG-free)

Member Data Documentation

class_weights

template<typename RETURNTYPE>

std::vector<RETURNTYPE> FST::Metric_Spec< RETURNTYPE >::class_weights

per-class cost weights, indexed by TRUE class; empty = unweighted. A weight scales the whole confusion-matrix ROW of its class, i.e. it is the relative importance of the samples of that class, not of the predictions made for it. Only the ratios matter - every classification metric here is a ratio of weighted counts, so scaling all weights by a common factor leaves the value unchanged. Example: weights (1,5) on two-class data make each class-1 sample count like five class-0 samples, so errors on class 1 become five times as expensive. Must hold at least classes entries when non-empty.

---

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

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