Problems.md

Input Data Interfaces

• Input Data Interfaces
  • For classification
    • Sample implementation
  • For regression

If you are going to train a classification or regression model, you will need to represent the input data as an object implementing one of the interfaces described below.

For classification

The input data set for training a classification model should be represented by an IProblem interface. The object should provide all the information for training. The main data is a set of vectors, and each vector contains the feature values for a single object. IProblem should also present the information about the classes and features to be used:

• GetClassCount — the number of classes
• GetFeatureCount — the number of features (that is, a single vector length)
• IsDiscreteFeature — indicates if the feature with the given index is discrete
• GetVectorCount — the number of vectors (and objects to be classified)
• GetClass — the class of the vector with the given number; classes are numbered from 0 to (GetClassCount - 1)
• GetVector — the vector with the given index
• GetMatrix — the whole training set as a matrix (of the GetFeatureCount * GetVectorCount size)
• GetVectorWeight — the vector weight

class NEOML_API IProblem : virtual public IObject {
public:
	virtual ~IProblem() = 0;

	// The number of classes
	virtual int GetClassCount() const = 0;

	// The number of features
	virtual int GetFeatureCount() const = 0;

	// Indicates if the feature is discrete
	virtual bool IsDiscreteFeature( int index ) const = 0;

	// The number of input vectors
	virtual int GetVectorCount() const = 0;

	// Get the number of class [0, GetClassCount()) for the given vector
	virtual int GetClass( int index ) const = 0;

	// Get a vector
	virtual CSparseFloatVectorDesc GetVector( int index ) const = 0;

	// The training set as a matrix
	virtual CSparseFloatMatrixDesc GetMatrix() const = 0;

	// The vector weight
	virtual double GetVectorWeight( int index ) const = 0;
};

Sample implementation

The library provides the CMemoryProblem class that implements the IProblem interface. It stores all data in memory.

For regression

The input data set for training a regression model should be represented by an object implementing an IRegressionProblem interface (if the function returns a number) or an IMultivariateRegressionProblem interface (if the function returns a vector). The base interface for both is IBaseRegressionProblem.

This object should contain all data for model training. The main data is a set of vectors, and each vector contains the feature values for a single object:

• GetFeatureCount — the number of features (that is, a single vector length)
• GetVectorCount — the number of vectors (and objects to be classified)
• GetVector — the vector with the given index
• GetMatrix — the whole training set as a matrix (of the GetFeatureCount * GetVectorCount size)
• GetVectorWeight — the vector weight
• GetValue — the value of the function on the given vector (a single number for IRegressionProblem, a vector for IMultivariateRegressionProblem);
• GetValueSize — (for IMultivariateRegressionProblem only) the length of the vector function value.

class IBaseRegressionProblem : virtual public IObject {
public:
	// The number of features
	virtual int GetFeatureCount() const = 0;

	// The number of input vectors
	virtual int GetVectorCount() const = 0;

	// Get a vector
	virtual CSparseFloatVectorDesc GetVector( int index ) const = 0;

	// The training set as a matrix
	virtual CSparseFloatMatrixDesc GetMatrix() const = 0;

	// The vector weight
	virtual double GetVectorWeight( int index ) const = 0;
};

// The data for training a regression model of a function with number value
class IRegressionProblem : public IBaseRegressionProblem {
public:
	// The function value on a vector
	virtual double GetValue( int index ) const = 0;
};

// The data for training a regression model of a function with vector value
class IMultivariateRegressionProblem : public IBaseRegressionProblem {
public:
	// The length of the function value
	virtual int GetValueSize() const = 0;
	// The function value on a vector
	virtual CFloatVector GetValue( int index ) const = 0;
};