FeatureSelectionNCARegression class

Superclasses:

Feature selection for regression using neighborhood component analysis (NCA)

Description

FeatureSelectionNCARegression contains the data, fitting information, feature weights, and other model parameters of a neighborhood component analysis (NCA) model. fsrnca learns the feature weights using a diagonal adaptation of NCA and returns an instance of FeatureSelectionNCARegression object. The function achieves feature selection by regularizing the feature weights.

Construction

Create a FeatureSelectionNCAClassification object using fsrnca.

Properties

expand all

Number of observations in the training data (X and Y) after removing NaN or Inf values, stored as a scalar.

Data Types: double

Model parameters used for training the model, stored as a structure.

You can access the fields of ModelParameters using dot notation.

For example, for a FeatureSelectionNCARegression object named mdl, you can access the LossFunction value using mdl.ModelParameters.LossFunction.

Data Types: struct

Regularization parameter used for training this model, stored as a scalar. For n observations, the best Lambda value that minimizes the generalization error of the NCA model is expected to be a multiple of 1/n.

Data Types: double

Name of the fitting method used to fit this model, stored as one of the following:

  • 'exact' — Perform fitting using all of the data.

  • 'none' — No fitting. Use this option to evaluate the generalization error of the NCA model using the initial feature weights supplied in the call to fsrnca.

  • 'average' — The software divides the data into partitions (subsets), fits each partition using the exact method, and returns the average of the feature weights. You can specify the number of partitions using the NumPartitions name-value pair argument.

Name of the solver used to fit this model, stored as one of the following:

  • 'lbfgs' — Limited memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) algorithm

  • 'sgd' — Stochastic gradient descent (SGD) algorithm

  • 'minibatch-lbfgs' — stochastic gradient descent with LBFGS algorithm applied to mini-batches

Relative convergence tolerance on the gradient norm for the 'lbfgs' and 'minibatch-lbfgs' solvers, stored as a positive scalar value.

Data Types: double

Maximum number of iterations for optimization, stored as a positive integer value.

Data Types: double

Maximum number of passes for 'sgd' and 'minibatch-lbfgs' solvers. Every pass processes all of the observations in the data.

Data Types: double

Initial learning rate for 'sgd' and 'minibatch-lbfgs' solvers. The learning rate decays over iterations starting at the value specified for InitialLearningRate.

Use the NumTuningIterations and TuningSubsetSize to control the automatic tuning of initial learning rate in the call to fsrnca.

Data Types: double

Verbosity level indicator, stored as a nonnegative integer. Possible values are:

  • 0 — No convergence summary

  • 1 — Convergence summary, including norm of gradient and objective function value

  • >1 — More convergence information, depending on the fitting algorithm. When you use the 'minibatch-lbfgs' solver and verbosity level > 1, the convergence information includes the iteration log from intermediate minibatch LBFGS fits.

Data Types: double

Initial feature weights, stored as a p-by-1 vector of positive real scalars, where p is the number of predictors in X.

Data Types: double

Feature weights, stored as a p-by-1 vector of real scalar values, where p is the number of predictors in X.

For 'FitMethod' equal to 'average', FeatureWeights is a p-by-m matrix, where m is the number of partitions specified via the 'NumPartitions' name-value pair argument in the call to fsrnca.

The absolute value of FeatureWeights(k) is a measure of the importance of predictor k. If FeatureWeights(k) is close to 0, then this indicates that predictor k does not influence the response in Y.

Data Types: double

Fit information, stored as a structure with the following fields.

Field NameMeaning
IterationIteration index
ObjectiveRegularized objective function for minimization
UnregularizedObjectiveUnregularized objective function for minimization
GradientGradient of regularized objective function for minimization
  • For classification, UnregularizedObjective represents the negative of the leave-one-out accuracy of the NCA classifier on the training data.

  • For regression, UnregularizedObjective represents the leave-one-out loss between the true response and the predicted response when using the NCA regression model.

  • For the 'lbfgs' solver, Gradient is the final gradient. For the 'sgd' and 'minibatch-lbfgs' solvers, Gradient is the final mini-batch gradient.

  • If FitMethod is 'average', then FitInfo is an m-by-1 structure array, where m is the number of partitions specified via the 'NumPartitions' name-value pair argument.

You can access the fields of FitInfo using dot notation. For example, for a FeatureSelectionNCARegressionobject named mdl, you can access the Objective field using mdl.FitInfo.Objective.

Data Types: struct

Predictor means, stored as a p-by-1 vector for standardized training data. In this case, the predict method centers predictor matrix X by subtracting the respective element of Mu from every column.

If data is not standardized during training, then Mu is empty.

Data Types: double

Predictor standard deviations, stored as a p-by-1 vector for standardized training data. In this case, the predict method scales predictor matrix X by dividing every column by the respective element of Sigma after centering the data using Mu.

If data is not standardized during training, then Sigma is empty.

Data Types: double

Predictor values used to train this model, stored as an n-by-p matrix. n is the number of observations and p is the number of predictor variables in the training data.

Data Types: double

Response values used to train this model, stored as a numeric vector of size n, where n is the number of observations.

Data Types: double

Observation weights used to train this model, stored as a numeric vector of size n. The sum of observation weights is n.

Data Types: double

Methods

loss Evaluate accuracy of learned feature weights on test data
predictPredict responses using neighborhood component analysis (NCA) regression model
refitRefit neighborhood component analysis (NCA) model for regression

Examples

collapse all

Load the sample data.

load imports-85

The first 15 columns contain the continuous predictor variables, whereas the 16th column contains the response variable, which is the price of a car. Define the variables for the neighborhood component analysis model.

Predictors = X(:,1:15);
Y = X(:,16);

Fit a neighborhood component analysis (NCA) model for regression to detect the relevant features.

mdl = fsrnca(Predictors,Y);

The returned NCA model, mdl, is a FeatureSelectionNCARegression object. This object stores information about the training data, model, and optimization. You can access the object properties, such as the feature weights, using dot notation.

Plot the feature weights.

figure()
plot(mdl.FeatureWeights,'ro')
xlabel('Feature Index')
ylabel('Feature Weight')
grid on

The weights of the irrelevant features are zero. The 'Verbose',1 option in the call to fsrnca displays the optimization information on the command line. You can also visualize the optimization process by plotting the objective function versus the iteration number.

figure()
plot(mdl.FitInfo.Iteration,mdl.FitInfo.Objective,'ro-')
grid on
xlabel('Iteration Number')
ylabel('Objective')

The ModelParameters property is a struct that contains more information about the model. You can access the fields of this property using dot notation. For example, see if the data was standardized or not.

mdl.ModelParameters.Standardize
ans = logical
   0

0 means that the data was not standardized before fitting the NCA model. You can standardize the predictors when they are on very different scales using the 'Standardize',1 name-value pair argument in the call to fsrnca .

Copy Semantics

Value. To learn how value classes affect copy operations, see Copying Objects (MATLAB).

See Also

| | |

Topics

Introduced in R2016b