predict

Classify observations using multiclass error-correcting output codes (ECOC) model

Description

example

label = predict(Mdl,X) returns a vector of predicted class labels (label) for the predictor data in the table or matrix X, based on the trained multiclass error-correcting output codes (ECOC) model Mdl. The trained ECOC model can be either full or compact.

example

label = predict(Mdl,X,Name,Value) uses additional options specified by one or more name-value pair arguments. For example, you can specify the posterior probability estimation method, decoding scheme, and verbosity level.

example

[label,NegLoss,PBScore] = predict(___) uses any of the input argument combinations in the previous syntaxes and additionally returns:

  • An array of negated average binary losses (NegLoss). For each observation in X, predict assigns the label of the class yielding the largest negated average binary loss (or, equivalently, the smallest average binary loss).

  • An array of positive-class scores (PBScore) for the observations classified by each binary learner.

example

[label,NegLoss,PBScore,Posterior] = predict(___) additionally returns posterior class probability estimates for the observations (Posterior).

To obtain posterior class probabilities, you must set 'FitPosterior',true when training the ECOC model using fitcecoc. Otherwise, predict throws an error.

Examples

collapse all

Load Fisher's iris data set. Specify the predictor data X, the response data Y, and the order of the classes in Y.

load fisheriris
X = meas;
Y = categorical(species);
classOrder = unique(Y);
rng(1); % For reproducibility

Train an ECOC model using SVM binary classifiers. Specify a 30% holdout sample, standardize the predictors using an SVM template, and specify the class order.

t = templateSVM('Standardize',true);
PMdl = fitcecoc(X,Y,'Holdout',0.30,'Learners',t,'ClassNames',classOrder);
Mdl = PMdl.Trained{1};           % Extract trained, compact classifier

PMdl is a ClassificationPartitionedECOC model. It has the property Trained, a 1-by-1 cell array containing the CompactClassificationECOC model that the software trained using the training set.

Predict the test-sample labels. Print a random subset of true and predicted labels.

testInds = test(PMdl.Partition);  % Extract the test indices
XTest = X(testInds,:);
YTest = Y(testInds,:);
labels = predict(Mdl,XTest);

idx = randsample(sum(testInds),10);
table(YTest(idx),labels(idx),...
    'VariableNames',{'TrueLabels','PredictedLabels'})
ans=10×2 table
    TrueLabels    PredictedLabels
    __________    _______________

    setosa          setosa       
    versicolor      virginica    
    setosa          setosa       
    virginica       virginica    
    versicolor      versicolor   
    setosa          setosa       
    virginica       virginica    
    virginica       virginica    
    setosa          setosa       
    setosa          setosa       

Mdl correctly labels all except one of the test-sample observations with indices idx.

Load Fisher's iris data set. Specify the predictor data X, the response data Y, and the order of the classes in Y.

load fisheriris
X = meas;
Y = categorical(species);
classOrder = unique(Y); % Class order
rng(1); % For reproducibility

Train an ECOC model using SVM binary classifiers and specify a 30% holdout sample. Standardize the predictors using an SVM template, and specify the class order.

t = templateSVM('Standardize',true);
PMdl = fitcecoc(X,Y,'Holdout',0.30,'Learners',t,'ClassNames',classOrder);
Mdl = PMdl.Trained{1};           % Extract trained, compact classifier

PMdl is a ClassificationPartitionedECOC model. It has the property Trained, a 1-by-1 cell array containing the CompactClassificationECOC model that the software trained using the training set.

SVM scores are signed distances from the observation to the decision boundary. Therefore, (-,) is the domain. Create a custom binary loss function that does the following:

  • Map the coding design matrix (M) and positive-class classification scores (s) for each learner to the binary loss for each observation.

  • Use linear loss.

  • Aggregate the binary learner loss using the median.

You can create a separate function for the binary loss function, and then save it on the MATLAB® path. Or, you can specify an anonymous binary loss function. In this case, create a function handle (customBL) to an anonymous binary loss function.

customBL = @(M,s)nanmedian(1 - bsxfun(@times,M,s),2)/2;

Predict test-sample labels and estimate the median binary loss per class. Print the median negative binary losses per class for a random set of 10 test-sample observations.

testInds = test(PMdl.Partition);  % Extract the test indices
XTest = X(testInds,:);
YTest = Y(testInds,:);
[label,NegLoss] = predict(Mdl,XTest,'BinaryLoss',customBL);

idx = randsample(sum(testInds),10);
classOrder
classOrder = 3x1 categorical array
     setosa 
     versicolor 
     virginica 

table(YTest(idx),label(idx),NegLoss(idx,:),'VariableNames',...
    {'TrueLabel','PredictedLabel','NegLoss'})
ans=10×3 table
    TrueLabel     PredictedLabel                 NegLoss              
    __________    ______________    __________________________________

    setosa          versicolor       0.18578       1.9878      -3.6736
    versicolor      virginica        -1.3316     -0.12355    -0.044843
    setosa          versicolor       0.13896        1.926       -3.565
    virginica       virginica        -1.5132     -0.38271      0.39588
    versicolor      versicolor      -0.87218      0.74736      -1.3752
    setosa          versicolor       0.48406       1.9977      -3.9818
    virginica       virginica        -1.9362     -0.67541       1.1117
    virginica       virginica        -1.5788     -0.83318      0.91194
    setosa          versicolor       0.51021       2.1212      -4.1314
    setosa          versicolor       0.36128       2.0596      -3.9209

The order of the columns corresponds to the elements of classOrder. The software predicts the label based on the maximum negated loss. The results indicate that the median of the linear losses might not perform as well as other losses.

Train an ECOC classifier using SVM binary learners. First predict the training-sample labels and class posterior probabilities. Then predict the maximum class posterior probability at each point in a grid. Visualize the results.

Load Fisher's iris data set. Specify the petal dimensions as the predictors and the species names as the response.

load fisheriris
X = meas(:,3:4);
Y = species;
rng(1); % For reproducibility

Create an SVM template. Standardize the predictors, and specify the Gaussian kernel.

t = templateSVM('Standardize',true,'KernelFunction','gaussian');

t is an SVM template. Most of its properties are empty. When the software trains the ECOC classifier, it sets the applicable properties to their default values.

Train the ECOC classifier using the SVM template. Transform classification scores to class posterior probabilities (which are returned by predict or resubPredict) using the 'FitPosterior' name-value pair argument. Specify the class order using the 'ClassNames' name-value pair argument. Display diagnostic messages during training by using the 'Verbose' name-value pair argument.

Mdl = fitcecoc(X,Y,'Learners',t,'FitPosterior',true,...
    'ClassNames',{'setosa','versicolor','virginica'},...
    'Verbose',2);
Training binary learner 1 (SVM) out of 3 with 50 negative and 50 positive observations.
Negative class indices: 2
Positive class indices: 1

Fitting posterior probabilities for learner 1 (SVM).
Training binary learner 2 (SVM) out of 3 with 50 negative and 50 positive observations.
Negative class indices: 3
Positive class indices: 1

Fitting posterior probabilities for learner 2 (SVM).
Training binary learner 3 (SVM) out of 3 with 50 negative and 50 positive observations.
Negative class indices: 3
Positive class indices: 2

Fitting posterior probabilities for learner 3 (SVM).

Mdl is a ClassificationECOC model. The same SVM template applies to each binary learner, but you can adjust options for each binary learner by passing in a cell vector of templates.

Predict the training-sample labels and class posterior probabilities. Display diagnostic messages during the computation of labels and class posterior probabilities by using the 'Verbose' name-value pair argument.

[label,~,~,Posterior] = resubPredict(Mdl,'Verbose',1);
Predictions from all learners have been computed.
Loss for all observations has been computed.
Computing posterior probabilities...
Mdl.BinaryLoss
ans = 
'quadratic'

The software assigns an observation to the class that yields the smallest average binary loss. Because all binary learners are computing posterior probabilities, the binary loss function is quadratic.

Display a random set of results.

idx = randsample(size(X,1),10,1);
Mdl.ClassNames
ans = 3x1 cell array
    {'setosa'    }
    {'versicolor'}
    {'virginica' }

table(Y(idx),label(idx),Posterior(idx,:),...
    'VariableNames',{'TrueLabel','PredLabel','Posterior'})
ans=10×3 table
      TrueLabel         PredLabel                     Posterior               
    ______________    ______________    ______________________________________

    {'virginica' }    {'virginica' }     0.0039316     0.0039864       0.99208
    {'virginica' }    {'virginica' }      0.017065      0.018261       0.96467
    {'virginica' }    {'virginica' }      0.014946      0.015854        0.9692
    {'versicolor'}    {'versicolor'}    2.2197e-14       0.87318       0.12682
    {'setosa'    }    {'setosa'    }         0.999    0.00025091     0.0007464
    {'versicolor'}    {'virginica' }    2.2195e-14      0.059423       0.94058
    {'versicolor'}    {'versicolor'}    2.2194e-14       0.97002      0.029983
    {'setosa'    }    {'setosa'    }         0.999    0.00024989    0.00074741
    {'versicolor'}    {'versicolor'}     0.0085637       0.98259     0.0088481
    {'setosa'    }    {'setosa'    }         0.999    0.00025012    0.00074719

The columns of Posterior correspond to the class order of Mdl.ClassNames.

Define a grid of values in the observed predictor space. Predict the posterior probabilities for each instance in the grid.

xMax = max(X);
xMin = min(X);

x1Pts = linspace(xMin(1),xMax(1));
x2Pts = linspace(xMin(2),xMax(2));
[x1Grid,x2Grid] = meshgrid(x1Pts,x2Pts);

[~,~,~,PosteriorRegion] = predict(Mdl,[x1Grid(:),x2Grid(:)]);

For each coordinate on the grid, plot the maximum class posterior probability among all classes.

contourf(x1Grid,x2Grid,...
        reshape(max(PosteriorRegion,[],2),size(x1Grid,1),size(x1Grid,2)));
h = colorbar;
h.YLabel.String = 'Maximum posterior';
h.YLabel.FontSize = 15;

hold on
gh = gscatter(X(:,1),X(:,2),Y,'krk','*xd',8);
gh(2).LineWidth = 2;
gh(3).LineWidth = 2;

title('Iris Petal Measurements and Maximum Posterior')
xlabel('Petal length (cm)')
ylabel('Petal width (cm)')
axis tight
legend(gh,'Location','NorthWest')
hold off

Train a multiclass ECOC model and estimate posterior probabilities using parallel computing.

Load the arrhythmia data set. Examine the response data Y, and determine the number of classes.

load arrhythmia
Y = categorical(Y);
tabulate(Y)
  Value    Count   Percent
      1      245     54.20%
      2       44      9.73%
      3       15      3.32%
      4       15      3.32%
      5       13      2.88%
      6       25      5.53%
      7        3      0.66%
      8        2      0.44%
      9        9      1.99%
     10       50     11.06%
     14        4      0.88%
     15        5      1.11%
     16       22      4.87%
K = numel(unique(Y));

Several classes are not represented in the data, and many of the other classes have low relative frequencies.

Specify an ensemble learning template that uses the GentleBoost method and 50 weak classification tree learners.

t = templateEnsemble('GentleBoost',50,'Tree');

t is a template object. Most of its properties are empty ([]). The software uses default values for all empty properties during training.

Because the response variable contains many classes, specify a sparse random coding design.

rng(1); % For reproducibility
Coding = designecoc(K,'sparserandom');

Train an ECOC model using parallel computing. Specify a 15% holdout sample, and fit posterior probabilities.

pool = parpool;                    % Invokes workers
Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 6).
options = statset('UseParallel',true);
PMdl = fitcecoc(X,Y,'Learner',t,'Options',options,'Coding',Coding,...
    'FitPosterior',true,'Holdout',0.15);
Mdl = PMdl.Trained{1};            % Extract trained, compact classifier

PMdl is a ClassificationPartitionedECOC model. It has the property Trained, a 1-by-1 cell array containing the CompactClassificationECOC model that the software trained using the training set.

The pool invokes six workers, although the number of workers might vary among systems.

Estimate posterior probabilities, and display the posterior probability of being classified as not having arrhythmia (class 1) given the data for a random set of test-sample observations.

testInds = test(PMdl.Partition);  % Extract the test indices
XTest = X(testInds,:);
YTest = Y(testInds,:);
[~,~,~,posterior] = predict(Mdl,XTest,'Options',options);

idx = randsample(sum(testInds),10);
table(idx,YTest(idx),posterior(idx,1),...
    'VariableNames',{'TestSampleIndex','TrueLabel','PosteriorNoArrhythmia'})
ans=10×3 table
    TestSampleIndex    TrueLabel    PosteriorNoArrhythmia
    _______________    _________    _____________________

          11              6                0.60631       
          41              4                0.23674       
          51              2                0.13802       
          33              10               0.43831       
          12              1                0.94332       
           8              1                0.97278       
          37              1                0.62807       
          24              10               0.96876       
          56              16               0.29375       
          30              1                0.64512       

Input Arguments

collapse all

Full or compact multiclass ECOC model, specified as a ClassificationECOC or CompactClassificationECOC model object.

To create a full or compact ECOC model, see ClassificationECOC or CompactClassificationECOC.

Predictor data to be classified, specified as a numeric matrix or table.

Each row of X corresponds to one observation, and each column corresponds to one variable.

  • For a numeric matrix:

    • The variables that constitute the columns of X must have the same order as the predictor variables that train Mdl.

    • If you train Mdl using a table (for example, Tbl), then X can be a numeric matrix if Tbl contains all numeric predictor variables. To treat numeric predictors in Tbl as categorical during training, identify categorical predictors using the CategoricalPredictors name-value pair argument of fitcecoc. 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 and 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 predictor variables that train Mdl (stored in Mdl.PredictorNames). However, the column order of X does not need to correspond to the column order of Tbl. Both 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 and the corresponding predictor variable names in X must be the same. To specify predictor names during training, see the PredictorNames name-value pair argument of fitcecoc. 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.

Note

If Mdl.BinaryLearners contains linear or kernel classification models (ClassificationLinear or ClassificationKernel model objects), then you cannot specify sample data in a table. Instead, pass a matrix of predictor data.

When training Mdl, assume that you set 'Standardize',true for a template object specified in the 'Learners' name-value pair argument of fitcecoc. In this case, for the corresponding binary learner j, the software standardizes the columns of the new predictor data using the corresponding means in Mdl.BinaryLearner{j}.Mu and standard deviations in Mdl.BinaryLearner{j}.Sigma.

Data Types: table | double | single

Name-Value Pair Arguments

Specify optional comma-separated pairs of Name,Value arguments. Name is the argument name and Value is the corresponding value. Name must appear inside quotes. You can specify several name and value pair arguments in any order as Name1,Value1,...,NameN,ValueN.

Example: predict(Mdl,X,'BinaryLoss','quadratic','Decoding','lossbased') specifies a quadratic binary learner loss function and a loss-based decoding scheme for aggregating the binary losses.

Binary learner loss function, specified as the comma-separated pair consisting of 'BinaryLoss' and a built-in loss function name or function handle.

  • This table describes the built-in functions, where yj is a class label for a particular binary learner (in the set {–1,1,0}), sj is the score for observation j, and g(yj,sj) is the binary loss formula.

    ValueDescriptionScore Domaing(yj,sj)
    'binodeviance'Binomial deviance(–∞,∞)log[1 + exp(–2yjsj)]/[2log(2)]
    'exponential'Exponential(–∞,∞)exp(–yjsj)/2
    'hamming'Hamming[0,1] or (–∞,∞)[1 – sign(yjsj)]/2
    'hinge'Hinge(–∞,∞)max(0,1 – yjsj)/2
    'linear'Linear(–∞,∞)(1 – yjsj)/2
    'logit'Logistic(–∞,∞)log[1 + exp(–yjsj)]/[2log(2)]
    'quadratic'Quadratic[0,1][1 – yj(2sj – 1)]2/2

    The software normalizes binary losses so that the loss is 0.5 when yj = 0. Also, the software calculates the mean binary loss for each class.

  • For a custom binary loss function, for example customFunction, specify its function handle 'BinaryLoss',@customFunction.

    customFunction has this form:

    bLoss = customFunction(M,s)
    where:

    • M is the K-by-L coding matrix stored in Mdl.CodingMatrix.

    • s is the 1-by-L row vector of classification scores.

    • bLoss is the classification loss. This scalar aggregates the binary losses for every learner in a particular class. For example, you can use the mean binary loss to aggregate the loss over the learners for each class.

    • K is the number of classes.

    • L is the number of binary learners.

    For an example of passing a custom binary loss function, see Predict Test-Sample Labels of ECOC Model Using Custom Binary Loss Function.

The default BinaryLoss value depends on the score ranges returned by the binary learners. This table describes some default BinaryLoss values based on the given assumptions.

AssumptionDefault Value
All binary learners are SVMs or either linear or kernel classification models of SVM learners.'hinge'
All binary learners are ensembles trained by AdaboostM1 or GentleBoost.'exponential'
All binary learners are ensembles trained by LogitBoost.'binodeviance'
All binary learners are linear or kernel classification models of logistic regression learners. Or, you specify to predict class posterior probabilities by setting 'FitPosterior',true in fitcecoc.'quadratic'

To check the default value, use dot notation to display the BinaryLoss property of the trained model at the command line.

Example: 'BinaryLoss','binodeviance'

Data Types: char | string | function_handle

Decoding scheme that aggregates the binary losses, specified as the comma-separated pair consisting of 'Decoding' and 'lossweighted' or 'lossbased'. For more information, see Binary Loss.

Example: 'Decoding','lossbased'

Number of random initial values for fitting posterior probabilities by Kullback-Leibler divergence minimization, specified as the comma-separated pair consisting of 'NumKLInitializations' and a nonnegative integer scalar.

If you do not request the fourth output argument (Posterior) and set 'PosteriorMethod','kl' (the default), then the software ignores the value of NumKLInitializations.

For more details, see Posterior Estimation Using Kullback-Leibler Divergence.

Example: 'NumKLInitializations',5

Data Types: single | double

Predictor data observation dimension, specified as the comma-separated pair consisting of 'ObservationsIn' and 'columns' or 'rows'. Mdl.BinaryLearners must contain ClassificationLinear models.

Note

If you orient your predictor matrix so that observations correspond to columns and specify 'ObservationsIn','columns', you can experience a significant reduction in execution time.

Estimation options, specified as the comma-separated pair consisting of 'Options' and a structure array returned by statset.

To invoke parallel computing:

  • You need a Parallel Computing Toolbox™ license.

  • Specify 'Options',statset('UseParallel',true).

Posterior probability estimation method, specified as the comma-separated pair consisting of 'PosteriorMethod' and 'kl' or 'qp'.

  • If PosteriorMethod is 'kl', then the software estimates multiclass posterior probabilities by minimizing the Kullback-Leibler divergence between the predicted and expected posterior probabilities returned by binary learners. For details, see Posterior Estimation Using Kullback-Leibler Divergence.

  • If PosteriorMethod is 'qp', then the software estimates multiclass posterior probabilities by solving a least-squares problem using quadratic programming. You need an Optimization Toolbox™ license to use this option. For details, see Posterior Estimation Using Quadratic Programming.

  • If you do not request the fourth output argument (Posterior), then the software ignores the value of PosteriorMethod.

Example: 'PosteriorMethod','qp'

Verbosity level, specified as the comma-separated pair consisting of 'Verbose' and 0 or 1. Verbose controls the number of diagnostic messages that the software displays in the Command Window.

If Verbose is 0, then the software does not display diagnostic messages. Otherwise, the software displays diagnostic messages.

Example: 'Verbose',1

Data Types: single | double

Output Arguments

collapse all

Predicted class labels, returned as a categorical, character, logical, or numeric array, or a cell array of character vectors. The software predicts the classification of an observation by assigning the observation to the class yielding the largest negated average binary loss (or, equivalently, the smallest average binary loss).

label has the same data type as the class labels used to train Mdl and has the same number of rows as X. (The software treats string arrays as cell arrays of character vectors.)

If Mdl.BinaryLearners contains ClassificationLinear models, then label is an m-by-L matrix, where m is the number of observations in X, and L is the number of regularization strengths in the linear classification models (numel(Mdl.BinaryLearners{1}.Lambda)). The value label(i,j) is the predicted label of observation i for the model trained using regularization strength Mdl.BinaryLearners{1}.Lambda(j).

Otherwise, label is a column vector of length m.

Negated average binary losses, returned as a numeric matrix or array.

  • If Mdl.BinaryLearners contains ClassificationLinear models, then NegLoss is an m-by-K-by-L array.

    • m is the number of observations in X.

    • K is the number of distinct classes in the training data (numel(Mdl.ClassNames)).

    • L is the number of regularization strengths in the linear classification models (numel(Mdl.BinaryLearners{1}.Lambda)).

    NegLoss(i,k,j) is the negated average binary loss for observation i, corresponding to class Mdl.ClassNames(k), for the model trained using regularization strength Mdl.BinaryLearners{1}.Lambda(j).

  • Otherwise, NegLoss is an m-by-K matrix.

Positive-class scores for each binary learner, returned as a numeric matrix or array.

  • If Mdl.BinaryLearners contains ClassificationLinear models, then PBScore is an m-by-B-by-L array.

    • m is the number of observations in X.

    • B is the number of binary learners (numel(Mdl.BinaryLearners)).

    • L is the number of regularization strengths in the linear classification models (numel(Mdl.BinaryLearners{1}.Lambda)).

    PBScore(i,b,j) is the positive-class score for observation i, using binary learner b, for the model trained using regularization strength Mdl.BinaryLearners{1}.Lambda(j).

  • Otherwise, PBScore is an m-by-B matrix.

Posterior class probabilities, returned as a numeric matrix or array.

  • If Mdl.BinaryLearners contains ClassificationLinear models, then Posterior is an m-by-K-by-L array. For dimension definitions, see NegLoss. Posterior(i,k,j) is the posterior probability that observation i comes from class Mdl.ClassNames(k), for the model trained using regularization strength Mdl.BinaryLearners{1}.Lambda(j).

  • Otherwise, Posterior is an m-by-K matrix.

More About

collapse all

Binary Loss

A binary loss is a function of the class and classification score that determines how well a binary learner classifies an observation into the class.

Suppose the following:

  • mkj is element (k,j) of the coding design matrix M (that is, the code corresponding to class k of binary learner j).

  • sj is the score of binary learner j for an observation.

  • g is the binary loss function.

  • k^ is the predicted class for the observation.

In loss-based decoding [Escalera et al.], the class producing the minimum sum of the binary losses over binary learners determines the predicted class of an observation, that is,

k^=argminkj=1L|mkj|g(mkj,sj).

In loss-weighted decoding [Escalera et al.], the class producing the minimum average of the binary losses over binary learners determines the predicted class of an observation, that is,

k^=argminkj=1L|mkj|g(mkj,sj)j=1L|mkj|.

Allwein et al. suggest that loss-weighted decoding improves classification accuracy by keeping loss values for all classes in the same dynamic range.

This table summarizes the supported loss functions, where yj is a class label for a particular binary learner (in the set {–1,1,0}), sj is the score for observation j, and g(yj,sj).

ValueDescriptionScore Domaing(yj,sj)
'binodeviance'Binomial deviance(–∞,∞)log[1 + exp(–2yjsj)]/[2log(2)]
'exponential'Exponential(–∞,∞)exp(–yjsj)/2
'hamming'Hamming[0,1] or (–∞,∞)[1 – sign(yjsj)]/2
'hinge'Hinge(–∞,∞)max(0,1 – yjsj)/2
'linear'Linear(–∞,∞)(1 – yjsj)/2
'logit'Logistic(–∞,∞)log[1 + exp(–yjsj)]/[2log(2)]
'quadratic'Quadratic[0,1][1 – yj(2sj – 1)]2/2

The software normalizes binary losses such that the loss is 0.5 when yj = 0, and aggregates using the average of the binary learners [Allwein et al.].

Do not confuse the binary loss with the overall classification loss (specified by the 'LossFun' name-value pair argument of the loss and predict object functions), which measures how well an ECOC classifier performs as a whole.

Algorithms

collapse all

The software can estimate class posterior probabilities by minimizing the Kullback-Leibler divergence or by using quadratic programming. For the following descriptions of the posterior estimation algorithms, assume that:

  • mkj is the element (k,j) of the coding design matrix M.

  • I is the indicator function.

  • p^k is the class posterior probability estimate for class k of an observation, k = 1,...,K.

  • rj is the positive-class posterior probability for binary learner j. That is, rj is the probability that binary learner j classifies an observation into the positive class, given the training data.

Posterior Estimation Using Kullback-Leibler Divergence

By default, the software minimizes the Kullback-Leibler divergence to estimate class posterior probabilities. The Kullback-Leibler divergence between the expected and observed positive-class posterior probabilities is

Δ(r,r^)=j=1Lwj[rjlogrjr^j+(1rj)log1rj1r^j],

where wj=Sjwi is the weight for binary learner j.

  • Sj is the set of observation indices on which binary learner j is trained.

  • wi is the weight of observation i.

The software minimizes the divergence iteratively. The first step is to choose initial values p^k(0);k=1,...,K for the class posterior probabilities.

  • If you do not specify 'NumKLIterations', then the software tries both sets of deterministic initial values described next, and selects the set that minimizes Δ.

    • p^k(0)=1/K;k=1,...,K.

    • p^k(0);k=1,...,K is the solution of the system

      M01p^(0)=r,

      where M01 is M with all mkj = –1 replaced with 0, and r is a vector of positive-class posterior probabilities returned by the L binary learners [Dietterich et al.]. The software uses lsqnonneg to solve the system.

  • If you specify 'NumKLIterations',c, where c is a natural number, then the software does the following to choose the set p^k(0);k=1,...,K, and selects the set that minimizes Δ.

    • The software tries both sets of deterministic initial values as described previously.

    • The software randomly generates c vectors of length K using rand, and then normalizes each vector to sum to 1.

At iteration t, the software completes these steps:

  1. Compute

    r^j(t)=k=1Kp^k(t)I(mkj=+1)k=1Kp^k(t)I(mkj=+1mkj=1).

  2. Estimate the next class posterior probability using

    p^k(t+1)=p^k(t)j=1Lwj[rjI(mkj=+1)+(1rj)I(mkj=1)]j=1Lwj[r^j(t)I(mkj=+1)+(1r^j(t))I(mkj=1)].

  3. Normalize p^k(t+1);k=1,...,K so that they sum to 1.

  4. Check for convergence.

For more details, see [Hastie et al.] and [Zadrozny].

Posterior Estimation Using Quadratic Programming

Posterior probability estimation using quadratic programming requires an Optimization Toolbox license. To estimate posterior probabilities for an observation using this method, the software completes these steps:

  1. Estimate the positive-class posterior probabilities, rj, for binary learners j = 1,...,L.

  2. Using the relationship between rj and p^k [Wu et al.], minimize

    j=1L[rjk=1Kp^kI(mkj=1)+(1rj)k=1Kp^kI(mkj=+1)]2

    with respect to p^k and the restrictions

    0p^k1kp^k=1.

    The software performs minimization using quadprog.

References

[1] Allwein, E., R. Schapire, and Y. Singer. “Reducing multiclass to binary: A unifying approach for margin classifiers.” Journal of Machine Learning Research. Vol. 1, 2000, pp. 113–141.

[2] Dietterich, T., and G. Bakiri. “Solving Multiclass Learning Problems Via Error-Correcting Output Codes.” Journal of Artificial Intelligence Research. Vol. 2, 1995, pp. 263–286.

[3] Escalera, S., O. Pujol, and P. Radeva. “On the decoding process in ternary error-correcting output codes.” IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 32, Issue 7, 2010, pp. 120–134.

[4] Escalera, S., O. Pujol, and P. Radeva. “Separability of ternary codes for sparse designs of error-correcting output codes.” Pattern Recognition. Vol. 30, Issue 3, 2009, pp. 285–297.

[5] Hastie, T., and R. Tibshirani. “Classification by Pairwise Coupling.” Annals of Statistics. Vol. 26, Issue 2, 1998, pp. 451–471.

[6] Wu, T. F., C. J. Lin, and R. Weng. “Probability Estimates for Multi-Class Classification by Pairwise Coupling.” Journal of Machine Learning Research. Vol. 5, 2004, pp. 975–1005.

[7] Zadrozny, B. “Reducing Multiclass to Binary by Coupling Probability Estimates.” NIPS 2001: Proceedings of Advances in Neural Information Processing Systems 14, 2001, pp. 1041–1048.

Extended Capabilities

Introduced in R2014b