# predict

**Class: **FeatureSelectionNCAClassification

Predict responses using neighborhood component analysis (NCA) classifier

## Description

## Input Arguments

`mdl`

— Neighborhood component analysis model for classification

`FeatureSelectionNCAClassification`

object

Neighborhood component analysis model for classification, specified
as a `FeatureSelectionNCAClassification`

object.

`X`

— Predictor variable values

table | *n*-by-*p* matrix

Predictor variable values, specified as a table or an
*n*-by-*p* matrix, where
*n* is the number of observations and
*p* is the number of predictor variables used to train
`mdl`

.

By default, each row of `X`

corresponds to one
observation, and each column corresponds to one variable.

For a numeric matrix:

The variables in the columns of

`X`

must have the same order as the predictor variables that trained`mdl`

.If you train

`mdl`

using a table (for example,`Tbl`

), and`Tbl`

contains only numeric predictor variables, then`X`

can be a numeric matrix. To treat numeric predictors in`Tbl`

as categorical during training, identify categorical predictors by using the`CategoricalPredictors`

name-value argument of`fscnca`

. If`Tbl`

contains heterogeneous predictor variables (for example, numeric and categorical data types), and`X`

is a numeric matrix, then`predict`

throws an error.

For a table:

`predict`

does not support multicolumn variables or cell arrays other than cell arrays of character vectors.If you train

`mdl`

using a table (for example,`Tbl`

), then all predictor variables in`X`

must have the same variable names and data types as the variables that trained`mdl`

(stored in`mdl.PredictorNames`

). However, the column order of`X`

does not need to correspond to the column order of`Tbl`

. Also,`Tbl`

and`X`

can contain additional variables (response variables, observation weights, and so on), but`predict`

ignores them.If you train

`mdl`

using a numeric matrix, then the predictor names in`mdl.PredictorNames`

must be the same as the corresponding predictor variable names in`X`

. To specify predictor names during training, use the`CategoricalPredictors`

name-value argument of`fscnca`

. All predictor variables in`X`

must be numeric vectors.`X`

can contain additional variables (response variables, observation weights, and so on), but`predict`

ignores them.

**Data Types: **`table`

| `single`

| `double`

## Output Arguments

`labels`

— Predicted class labels

categorical vector | logical vector | numeric vector | cell array of character vectors | character array

Predicted class labels corresponding to the rows of `X`

, returned as a
categorical, logical, or numeric vector, a cell array of character vectors
of length *n*, or a character array with
*n* rows. *n* is the number of
observations. The type of `labels`

is the same as for
`ResponseName`

or `Y`

used in
training.

`postprobs`

— Posterior probabilities

*n*-by-*c* matrix

Posterior probabilities, returned as an *n*-by-*c* matrix,
where *n* is the number of observations and *c* is
the number of classes. A posterior probability, `postprobs(i,:)`

,
represents the membership of an observation in `X(i,:)`

in
classes 1 through *c*.

`classnames`

— Class names

cell array of character vectors

Class names corresponding to posterior probabilities, returned
as a cell array of character vectors. Each character vector is the
class name corresponding to a column of `postprobs`

.

## Examples

### Tune NCA Model for Classification

Load the sample data.

`load('twodimclassdata.mat');`

This data set is simulated using the scheme described in [1]. This is a two-class classification problem in two dimensions. Data from the first class (class –1) are drawn from two bivariate normal distributions $$N({\mu}_{1},\Sigma )$$ or $$N({\mu}_{2},\Sigma )$$ with equal probability, where $${\mu}_{1}=[-0.75,-1.5]$$, $${\mu}_{2}=[0.75,1.5]$$, and $$\Sigma ={I}_{2}$$. Similarly, data from the second class (class 1) are drawn from two bivariate normal distributions $$N({\mu}_{3},\Sigma )$$ or $$N({\mu}_{4},\Sigma )$$ with equal probability, where $${\mu}_{3}=[1.5,-1.5]$$, $${\mu}_{4}=[-1.5,1.5]$$, and $$\Sigma ={I}_{2}$$. The normal distribution parameters used to create this data set result in tighter clusters in data than the data used in [1].

Create a scatter plot of the data grouped by the class.

figure gscatter(X(:,1),X(:,2),y) xlabel('x1') ylabel('x2')

Add 100 irrelevant features to $$X$$. First generate data from a Normal distribution with a mean of 0 and a variance of 20.

```
n = size(X,1);
rng('default')
XwithBadFeatures = [X,randn(n,100)*sqrt(20)];
```

Normalize the data so that all points are between 0 and 1.

XwithBadFeatures = (XwithBadFeatures-min(XwithBadFeatures,[],1))./range(XwithBadFeatures,1); X = XwithBadFeatures;

Fit a neighborhood component analysis (NCA) model to the data using the default `Lambda`

(regularization parameter, $$\lambda $$) value. Use the LBFGS solver and display the convergence information.

ncaMdl = fscnca(X,y,'FitMethod','exact','Verbose',1, ... 'Solver','lbfgs');

o Solver = LBFGS, HessianHistorySize = 15, LineSearchMethod = weakwolfe |====================================================================================================| | ITER | FUN VALUE | NORM GRAD | NORM STEP | CURV | GAMMA | ALPHA | ACCEPT | |====================================================================================================| | 0 | 9.519258e-03 | 1.494e-02 | 0.000e+00 | | 4.015e+01 | 0.000e+00 | YES | | 1 | -3.093574e-01 | 7.186e-03 | 4.018e+00 | OK | 8.956e+01 | 1.000e+00 | YES | | 2 | -4.809455e-01 | 4.444e-03 | 7.123e+00 | OK | 9.943e+01 | 1.000e+00 | YES | | 3 | -4.938877e-01 | 3.544e-03 | 1.464e+00 | OK | 9.366e+01 | 1.000e+00 | YES | | 4 | -4.964759e-01 | 2.901e-03 | 6.084e-01 | OK | 1.554e+02 | 1.000e+00 | YES | | 5 | -4.972077e-01 | 1.323e-03 | 6.129e-01 | OK | 1.195e+02 | 5.000e-01 | YES | | 6 | -4.974743e-01 | 1.569e-04 | 2.155e-01 | OK | 1.003e+02 | 1.000e+00 | YES | | 7 | -4.974868e-01 | 3.844e-05 | 4.161e-02 | OK | 9.835e+01 | 1.000e+00 | YES | | 8 | -4.974874e-01 | 1.417e-05 | 1.073e-02 | OK | 1.043e+02 | 1.000e+00 | YES | | 9 | -4.974874e-01 | 4.893e-06 | 1.781e-03 | OK | 1.530e+02 | 1.000e+00 | YES | | 10 | -4.974874e-01 | 9.404e-08 | 8.947e-04 | OK | 1.670e+02 | 1.000e+00 | YES | Infinity norm of the final gradient = 9.404e-08 Two norm of the final step = 8.947e-04, TolX = 1.000e-06 Relative infinity norm of the final gradient = 9.404e-08, TolFun = 1.000e-06 EXIT: Local minimum found.

Plot the feature weights. The weights of the irrelevant features should be very close to zero.

figure semilogx(ncaMdl.FeatureWeights,'ro') xlabel('Feature index') ylabel('Feature weight') grid on

Predict the classes using the NCA model and compute the confusion matrix.

ypred = predict(ncaMdl,X); confusionchart(y,ypred)

Confusion matrix shows that 40 of the data that are in class –1 are predicted as belonging to class –1. 60 of the data from class –1 are predicted to be in class 1. Similarly, 94 of the data from class 1 are predicted to be from class 1 and 6 of them are predicted to be from class –1. The prediction accuracy for class –1 is not good.

All weights are very close to zero, which indicates that the value of $$\lambda $$ used in training the model is too large. When $$\lambda \to \infty $$, all features weights approach to zero. Hence, it is important to tune the regularization parameter in most cases to detect the relevant features.

Use five-fold cross-validation to tune $$\lambda $$ for feature selection by using `fscnca`

. Tuning $$\lambda $$ means finding the $$\lambda $$ value that will produce the minimum classification loss. To tune $$\lambda $$ using cross-validation:

1. Partition the data into five folds. For each fold, `cvpartition`

assigns four-fifths of the data as a training set and one-fifth of the data as a test set. Again for each fold, `cvpartition`

creates a stratified partition, where each partition has roughly the same proportion of classes.

```
cvp = cvpartition(y,'kfold',5);
numtestsets = cvp.NumTestSets;
lambdavalues = linspace(0,2,20)/length(y);
lossvalues = zeros(length(lambdavalues),numtestsets);
```

2. Train the neighborhood component analysis (NCA) model for each $$\lambda $$ value using the training set in each fold.

3. Compute the classification loss for the corresponding test set in the fold using the NCA model. Record the loss value.

4. Repeat this process for all folds and all $$\lambda $$ values.

for i = 1:length(lambdavalues) for k = 1:numtestsets % Extract the training set from the partition object Xtrain = X(cvp.training(k),:); ytrain = y(cvp.training(k),:); % Extract the test set from the partition object Xtest = X(cvp.test(k),:); ytest = y(cvp.test(k),:); % Train an NCA model for classification using the training set ncaMdl = fscnca(Xtrain,ytrain,'FitMethod','exact', ... 'Solver','lbfgs','Lambda',lambdavalues(i)); % Compute the classification loss for the test set using the NCA % model lossvalues(i,k) = loss(ncaMdl,Xtest,ytest, ... 'LossFunction','quadratic'); end end

Plot the average loss values of the folds versus the $$\lambda $$ values. If the $$\lambda $$ value that corresponds to the minimum loss falls on the boundary of the tested $$\lambda $$ values, the range of $$\lambda $$ values should be reconsidered.

figure plot(lambdavalues,mean(lossvalues,2),'ro-') xlabel('Lambda values') ylabel('Loss values') grid on

Find the $$\lambda $$ value that corresponds to the minimum average loss.

[~,idx] = min(mean(lossvalues,2)); % Find the index bestlambda = lambdavalues(idx) % Find the best lambda value

bestlambda = 0.0037

Fit the NCA model to all of the data using the best $$\lambda $$ value. Use the LBFGS solver and display the convergence information.

ncaMdl = fscnca(X,y,'FitMethod','exact','Verbose',1, ... 'Solver','lbfgs','Lambda',bestlambda);

o Solver = LBFGS, HessianHistorySize = 15, LineSearchMethod = weakwolfe |====================================================================================================| | ITER | FUN VALUE | NORM GRAD | NORM STEP | CURV | GAMMA | ALPHA | ACCEPT | |====================================================================================================| | 0 | -1.246913e-01 | 1.231e-02 | 0.000e+00 | | 4.873e+01 | 0.000e+00 | YES | | 1 | -3.411330e-01 | 5.717e-03 | 3.618e+00 | OK | 1.068e+02 | 1.000e+00 | YES | | 2 | -5.226111e-01 | 3.763e-02 | 8.252e+00 | OK | 7.825e+01 | 1.000e+00 | YES | | 3 | -5.817731e-01 | 8.496e-03 | 2.340e+00 | OK | 5.591e+01 | 5.000e-01 | YES | | 4 | -6.132632e-01 | 6.863e-03 | 2.526e+00 | OK | 8.228e+01 | 1.000e+00 | YES | | 5 | -6.135264e-01 | 9.373e-03 | 7.341e-01 | OK | 3.244e+01 | 1.000e+00 | YES | | 6 | -6.147894e-01 | 1.182e-03 | 2.933e-01 | OK | 2.447e+01 | 1.000e+00 | YES | | 7 | -6.148714e-01 | 6.392e-04 | 6.688e-02 | OK | 3.195e+01 | 1.000e+00 | YES | | 8 | -6.149524e-01 | 6.521e-04 | 9.934e-02 | OK | 1.236e+02 | 1.000e+00 | YES | | 9 | -6.149972e-01 | 1.154e-04 | 1.191e-01 | OK | 1.171e+02 | 1.000e+00 | YES | | 10 | -6.149990e-01 | 2.922e-05 | 1.983e-02 | OK | 7.365e+01 | 1.000e+00 | YES | | 11 | -6.149993e-01 | 1.556e-05 | 8.354e-03 | OK | 1.288e+02 | 1.000e+00 | YES | | 12 | -6.149994e-01 | 1.147e-05 | 7.256e-03 | OK | 2.332e+02 | 1.000e+00 | YES | | 13 | -6.149995e-01 | 1.040e-05 | 6.781e-03 | OK | 2.287e+02 | 1.000e+00 | YES | | 14 | -6.149996e-01 | 9.015e-06 | 6.265e-03 | OK | 9.974e+01 | 1.000e+00 | YES | | 15 | -6.149996e-01 | 7.763e-06 | 5.206e-03 | OK | 2.919e+02 | 1.000e+00 | YES | | 16 | -6.149997e-01 | 8.374e-06 | 1.679e-02 | OK | 6.878e+02 | 1.000e+00 | YES | | 17 | -6.149997e-01 | 9.387e-06 | 9.542e-03 | OK | 1.284e+02 | 5.000e-01 | YES | | 18 | -6.149997e-01 | 3.250e-06 | 5.114e-03 | OK | 1.225e+02 | 1.000e+00 | YES | | 19 | -6.149997e-01 | 1.574e-06 | 1.275e-03 | OK | 1.808e+02 | 1.000e+00 | YES | |====================================================================================================| | ITER | FUN VALUE | NORM GRAD | NORM STEP | CURV | GAMMA | ALPHA | ACCEPT | |====================================================================================================| | 20 | -6.149997e-01 | 5.764e-07 | 6.765e-04 | OK | 2.905e+02 | 1.000e+00 | YES | Infinity norm of the final gradient = 5.764e-07 Two norm of the final step = 6.765e-04, TolX = 1.000e-06 Relative infinity norm of the final gradient = 5.764e-07, TolFun = 1.000e-06 EXIT: Local minimum found.

Plot the feature weights.

figure semilogx(ncaMdl.FeatureWeights,'ro') xlabel('Feature index') ylabel('Feature weight') grid on

`fscnca`

correctly figures out that the first two features are relevant and that the rest are not. The first two features are not individually informative, but when taken together result in an accurate classification model.

Predict the classes using the new model and compute the accuracy.

ypred = predict(ncaMdl,X); confusionchart(y,ypred)

Confusion matrix shows that prediction accuracy for class –1 has improved. 88 of the data from class –1 are predicted to be from –1, and 12 of them are predicted to be from class 1. 92 of the data from class 1 are predicted to be from class 1 and 8 of them are predicted to be from class –1.

**References**

[1] Yang, W., K. Wang, W. Zuo. "Neighborhood Component Feature Selection for High-Dimensional Data." *Journal of Computers*. Vol. 7, Number 1, January, 2012.

## Version History

**Introduced in R2016b**

## See Also

`FeatureSelectionNCAClassification`

| `fscnca`

| `loss`

| `refit`

| `selectFeatures`

## Open Example

You have a modified version of this example. Do you want to open this example with your edits?

## MATLAB Command

You clicked a link that corresponds to this MATLAB command:

Run the command by entering it in the MATLAB Command Window. Web browsers do not support MATLAB commands.

Select a Web Site

Choose a web site to get translated content where available and see local events and offers. Based on your location, we recommend that you select: .

You can also select a web site from the following list:

## How to Get Best Site Performance

Select the China site (in Chinese or English) for best site performance. Other MathWorks country sites are not optimized for visits from your location.

### Americas

- América Latina (Español)
- Canada (English)
- United States (English)

### Europe

- Belgium (English)
- Denmark (English)
- Deutschland (Deutsch)
- España (Español)
- Finland (English)
- France (Français)
- Ireland (English)
- Italia (Italiano)
- Luxembourg (English)

- Netherlands (English)
- Norway (English)
- Österreich (Deutsch)
- Portugal (English)
- Sweden (English)
- Switzerland
- United Kingdom (English)