Fn_sffs

Functions
double	ecocpak::criterion_fqmi (const vector< ClassData > &classes_vector, const ucolvec &indexPtr, const u32 curSubsetSize, const double sigma)
double	ecocpak::criterion_fldr (const vector< ClassData > &classes_vector, const ucolvec &indexPtr, const u32 curSubsetSize)
double	ecocpak::criterion (const vector< ClassData > &class_vector, const ucolvec &indexPtr, const u32 curSubSetSize, const u32 criterion_option, const double sigma)
ucolvec	ecocpak::sffs (const vector< ClassData > &classes_vector, const u32 crit)
double	criterion_fqmi (const vector< ClassData > &classes_vector, const ucolvec &indexPtr, const u32 curSubsetSize, const double sigma)
double	criterion_fldr (const vector< ClassData > &classes_vector, const ucolvec &indexPtr, const u32 curSubsetSize)
double	criterion (const vector< ClassData > &class_vector, const ucolvec &indexPtr, const u32 curSubSetSize, const u32 criterion_option, const double sigma)
ucolvec	sffs (const vector< ClassData > &classes_vector, const u32 crit)

---

Function Documentation

double ecocpak::criterion	(	const vector< ClassData > &	class_vector,
		const ucolvec &	indexPtr,
		const u32	curSubSetSize,
		const u32	criterion_option,
		const double	sigma
	)

• Criterion evaluation function for sffs() function.

• Input Arguments:
  • class_vector : Vector of classes.
  • indexPtr : Current set bin.
  • curSubSetSize : First position of indexPtr that represent current best subset.
  • criterion_option : Evaluation criterion.
  • sigma : Parameter used in the fqmi computation.
• Output Arguments:
  • Void.

• Return Argument:
  • Criterion value

double criterion	(	const vector< ClassData > &	class_vector,
		const ucolvec &	indexPtr,
		const u32	curSubSetSize,
		const u32	criterion_option,
		const double	sigma
	)

• Criterion evaluation function for sffs() function.

• Input Arguments:
  • class_vector : Vector of classes.
  • indexPtr : Current set bin.
  • curSubSetSize : First position of indexPtr that represent current best subset.
  • criterion_option : Evaluation criterion.
  • sigma : Parameter used in the fqmi computation.
• Output Arguments:
  • Void.

• Return Argument:
  • Criterion value

double ecocpak::criterion_fldr	(	const vector< ClassData > &	classes_vector,
		const ucolvec &	indexPtr,
		const u32	curSubsetSize
	)

• Fisher discriminant as a criterion function for SFFS algorithm.

• Criterion function for SFFS algorithm.

• Actually interface function for fldr().

• Input arguments:
  • classes_vector : Vector with classes.
  • indexPtr : Array of labels.
  • curSubSetSize : Current set's size.

• Output argument:
  • Scalar which denotes criterion value.

double criterion_fldr	(	const vector< ClassData > &	classes_vector,
		const ucolvec &	indexPtr,
		const u32	curSubsetSize
	)

• Fisher discriminant as a criterion function for SFFS algorithm.

• Criterion function for SFFS algorithm.

• Actually interface function for fldr().

• Input arguments:
  • classes_vector : Vector with classes.
  • indexPtr : Array of labels.
  • curSubSetSize : Current set's size.

• Output argument:
  • Scalar which denotes criterion value.

double criterion_fqmi	(	const vector< ClassData > &	classes_vector,
		const ucolvec &	indexPtr,
		const u32	curSubsetSize,
		const double	sigma
	)

• Fast Quadratic Mutual Information between classes and their respective labels.

• Criterion function for SFFS algorithm.

• Standard Criterion.

• Actually interface function for fqmi().

• Input Arguments:
  • classes_vector : Vector with classes.
  • indexPtr : Array of labels.
  • curSubSetSize : Current set's size.

• Output Arguments:
  • Void.

• Return Argument:
  • Scalar which denotes criterion value.

double ecocpak::criterion_fqmi	(	const vector< ClassData > &	classes_vector,
		const ucolvec &	indexPtr,
		const u32	curSubsetSize,
		const double	sigma
	)

• Fast Quadratic Mutual Information between classes and their respective labels.

• Criterion function for SFFS algorithm.

• Standard Criterion.

• Actually interface function for fqmi().

• Input Arguments:
  • classes_vector : Vector with classes.
  • indexPtr : Array of labels.
  • curSubSetSize : Current set's size.

• Output Arguments:
  • Void.

• Return Argument:
  • Scalar which denotes criterion value.

ucolvec sffs	(	const vector< ClassData > &	classes_vector,
		const u32	crit
	)

• This is a C++ listing of Classical Forward Floating Search Method. The original code was taken from Feature Selection Toolbox package programmed by Petr Somol, which was mainly written in C style and transformed in C++ style by Nikolaos Arvanitopoulos and Dimitrios Bouzas.

• To use the code you would probably have to rewrite/modify following parts:
  • criterion procedure "criterion()" should be replaced by calling of your criterion function, together with possible parameter passing.

• meaning of constants and some identifiers you may need to set for calling the procedure (local variables are described inside the procedure):
  • int n - full feature set size
  • int d - desired feature subset size
  • int delta - floating search stopping constraint
    • delta = 0 : full search (the procedure will not stop after reaching dimension d; SFFS will reach n, SFBS will reach 0).
    • delta > 0 : SFFS will float until dimension d+delta is reached, SFBS will float until d-delta is reached.
    • delta = 1 : The value of delta limit is computed during the course of the algorithm by averaging the length of backtracking.
    • delta = 2 : The value of delta limit is set during the course of the algorithm as the maximum length of backtracking.
  • int detail : textual output detail level (use NOTHING=0 or STANDARD=2).

• A non-redundand coding of subset configurations is used. For easier understanding imagine our goal is to find a subset by exhaustive search, when d = 5 and n = 12.

• Initial configuration is 111110000000 (actually stored in "bin" field).

• In every step:
  • find the leftmost block of 1's.
  • shift its rightmost 1 to the right.
  • shift the rest of this block to the left boundary (if it is not there).

• This algorithm generates increasingly all binary representations of subsets.

• In context of floating search this algorithm is used for computation of generalized steps (in case of simple SFFS and simple SFBS only single-feature configurations are tested).

• For purposes of floating search more identifiers than 0 and 1 are used to identify temporarily freezed features etc. (2,-1).

• Because of possibility to exchange meanings of 0 and 1 it was suitable to incorporate both forward and backward versions of floating search into one procedure.

• Input Arguments:
  • classes_vector : Vector of data classes.
  • crit : Criterion option.

• Output Arguments:
  • Void.

• Output Argument:
  • Vector of best subset's class indices.

ucolvec ecocpak::sffs	(	const vector< ClassData > &	classes_vector,
		const u32	crit
	)

• This is a C++ listing of Classical Forward Floating Search Method. The original code was taken from Feature Selection Toolbox package programmed by Petr Somol, which was mainly written in C style and transformed in C++ style by Nikolaos Arvanitopoulos and Dimitrios Bouzas.

• To use the code you would probably have to rewrite/modify following parts:
  • criterion procedure "criterion()" should be replaced by calling of your criterion function, together with possible parameter passing.

• meaning of constants and some identifiers you may need to set for calling the procedure (local variables are described inside the procedure):
  • int n - full feature set size
  • int d - desired feature subset size
  • int delta - floating search stopping constraint
    • delta = 0 : full search (the procedure will not stop after reaching dimension d; SFFS will reach n, SFBS will reach 0).
    • delta > 0 : SFFS will float until dimension d+delta is reached, SFBS will float until d-delta is reached.
    • delta = 1 : The value of delta limit is computed during the course of the algorithm by averaging the length of backtracking.
    • delta = 2 : The value of delta limit is set during the course of the algorithm as the maximum length of backtracking.
  • int detail : textual output detail level (use NOTHING=0 or STANDARD=2).

• A non-redundand coding of subset configurations is used. For easier understanding imagine our goal is to find a subset by exhaustive search, when d = 5 and n = 12.

• Initial configuration is 111110000000 (actually stored in "bin" field).

• In every step:
  • find the leftmost block of 1's.
  • shift its rightmost 1 to the right.
  • shift the rest of this block to the left boundary (if it is not there).

• This algorithm generates increasingly all binary representations of subsets.

• In context of floating search this algorithm is used for computation of generalized steps (in case of simple SFFS and simple SFBS only single-feature configurations are tested).

• For purposes of floating search more identifiers than 0 and 1 are used to identify temporarily freezed features etc. (2,-1).

• Because of possibility to exchange meanings of 0 and 1 it was suitable to incorporate both forward and backward versions of floating search into one procedure.

• Input Arguments:
  • classes_vector : Vector of data classes.
  • crit : Criterion option.

• Output Arguments:
  • Void.

• Output Argument:
  • Vector of best subset's class indices.