MaximalInformationCoefficientSelector<T>

Models & Types · AiDotNet.Preprocessing.FeatureSelection.Nonlinear

Maximal Information Coefficient (MIC) feature selection (Reshef et al. 2011, “Detecting Novel Associations in Large Data Sets”, Science 334).

Inheritance: TransformerBase<T, Matrix<T>, Matrix<T>> → MaximalInformationCoefficientSelector<T>

Implements: IFixedCountFeatureSelector<T>, ISupervisedFeatureSelector<T>

For Beginners

MIC is like correlation but can detect any type of relationship, not just linear ones. It scores relationships on a scale of 0 to 1, where 1 means a perfect (possibly complex) relationship exists.

How It Works

Scores every feature by its MIC with the target and keeps the highest-scoring ones. MIC is the largest normalized mutual information I(X; Y) / log(min(x, y)) over the x-by-y grids with x y <= B(n), where B(n) = n^alpha. It is computed with the paper’s ApproxMaxMI heuristic exactly as minepy does (equipartitioned rows, at most c x superclumps per column, the OptimizeXAxis dynamic program, both axis orders).

Constructors

Constructor Summary
MaximalInformationCoefficientSelector(int nFeaturesToSelect, MaximalInformationOptions options, int[] columnIndices) Creates a MIC feature selector.

Properties

Property Summary
Vector<T> MICValues { get; } The MIC of every feature with the target.
bool SupportsInverseTransform { get; }
double Alpha { get; } The exponent of the grid budget B(n) = n^alpha (or B itself when it is at least 4).
double C { get; } The clump factor: the x axis is split into at most c times as many superclumps as columns.
int NFeaturesToSelect { get; }
int[] SelectedIndices { get; }

Methods

Method Summary
Matrix<T> FitTransform(Matrix<T> data, Vector<T> target)
bool[] GetSupportMask()
string[] GetFeatureNamesOut(string[] inputFeatureNames)
void Fit(Matrix<T> data, Vector<T> target)