knnFit¶

Purpose¶

Creates a K-D tree model from training data for efficient KNN predictions.

Format¶

mdl = knnFit(y, X, k)¶

Parameters:

y (Nx1 vector or string array) – The dependent, or target, variable.
X (NxP matrix) – The independent, or features, variables.
k (Scalar) – The number of neighbors.

Returns:

mdl (struct) –

An instance of a knnModel structure. For an instance named mdl, the members will be:

mdl.opaqueModel	Column vector, containing the K-D tree in opaque form.
mdl.classIndices	Px1 matrix, where P is the number of classes in the target vector y.
mdl.classNames	Px1 string array, where P is the number of classes in the target vector y, containing the class names if the target vector was a string array.
mdl.k	Scalar, the number of neighbors to search.

Remarks¶

The KD-tree algorithm is an optimized approximate nearest neighbors algorithm. It provides high computational performance and high accuracy, but may not always return the exact nearest neighbors.

Examples¶

new;
library gml;

// Set seed for repeatable train/test sampling
rndseed 423432;

/*
** Load data and prepare
*/
// Get file name with full path
fname = getGAUSSHome("pkgs/gml/examples/iris.csv");

// Get predictors
X = loadd(fname, ". -Species");

// Load labels
species = loadd(fname, "Species");

// Split data into (70%) train and (30%) test sets
{ y_train, y_test, X_train, X_test } = trainTestSplit(species, X, 0.7);

/*
** Train the model
*/
// Specify number of neighbors
k = 3;

struct knnModel mdl;
mdl = knnFit(y_train, X_train, k);

/*
** Predictions on the test set
*/
y_hat = knnClassify(mdl, X_test);


/*
** Model assessment
*/
call classificationMetrics(y_test, y_hat);

The above code will print the following output:

===========================================================================
Model:                          KNN     Target Variable:            Species
Number observations:            105     Number features:                  4
Num. Neighbors:                   3     Number of Classes:                3
===========================================================================

KNN Classification Prediction Frequencies:
=============================================

     Label      Count   Total %    Cum. %
    setosa         14     31.11     31.11
versicolor         19     42.22     73.33
 virginica         12     26.67       100
     Total         45       100

=============================================

Observed Test Data Frequencies:
=============================================

     Label      Count   Total %    Cum. %
    setosa         14     31.11     31.11
versicolor         19     42.22     73.33
 virginica         12     26.67       100
     Total         45       100

=============================================

===================================================
                             Classification metrics
===================================================
       Class   Precision  Recall  F1-score  Support

      setosa        1.00    1.00      1.00       14
  versicolor        0.95    0.95      0.95       19
   virginica        0.92    0.92      0.92       12

   Macro avg        0.95    0.95      0.95       45
Weighted avg        0.96    0.96      0.96       45

    Accuracy                          0.96       45