Probabilistic prediction with liblinear (java), with direct usage of the classifier inside the code

Question

Consider the following usage of liblinear (http://liblinear.bwaldvogel.de/):

    double C = 1.0; // cost of constraints violation
    double eps = 0.01; // stopping criteria
    Parameter param = new Parameter(SolverType.L2R_L2LOSS_SVC, C, eps);
    Problem problem = new Problem();
    double[] GROUPS_ARRAY = {1, 0, 0, 0};
    problem.y = GROUPS_ARRAY;

    int NUM_OF_TS_EXAMPLES = 4;
    problem.l = NUM_OF_TS_EXAMPLES;
     problem.n = 2;

    FeatureNode[] instance1 = { new FeatureNode(1, 1), new FeatureNode(2, 1) };
    FeatureNode[] instance2 = { new FeatureNode(1, -1), new FeatureNode(2, 1) };
    FeatureNode[] instance3 = { new FeatureNode(1, -1), new FeatureNode(2, -1) };
    FeatureNode[] instance4 = { new FeatureNode(1, 1), new FeatureNode(2, -1) };

    FeatureNode[] instance5 = { new FeatureNode(1, 1), new FeatureNode(2, -0.1) };
    FeatureNode[] instance6 = { new FeatureNode(1, -0.1), new FeatureNode(2, 1) };
    FeatureNode[] instance7 = { new FeatureNode(1, -0.1), new FeatureNode(2, -0.1) };

    FeatureNode[][] testSetWithUnknown = {
            instance5,
            instance6, 
            instance7
        };

    FeatureNode[][] trainingSetWithUnknown = {
            instance1,
            instance2, 
            instance3, 
            instance4
        };

    problem.x = trainingSetWithUnknown;

    Model m = Linear.train(problem, param); 

    for( int i = 0; i < trainingSetWithUnknown.length; i++)
        System.out.println(" Train.instance =  " + i + " =>  " + Linear.predict(m, trainingSetWithUnknown[i]) ); 
    System.out.println("---------------------"); 
    for( int i = 0; i < testSetWithUnknown.length; i++)
        System.out.println(" Test.instance =  " + i + " =>  " + Linear.predict(m, testSetWithUnknown[i]) ); 

Here is the output :

iter  1 act 1.778e+00 pre 1.778e+00 delta 6.285e-01 f 4.000e+00 |g| 5.657e+00 CG   1
 Train.instance =  0 =>  1.0
 Train.instance =  1 =>  0.0
 Train.instance =  2 =>  0.0
 Train.instance =  3 =>  0.0
---------------------
 Test.instance =  0 =>  1.0
 Test.instance =  1 =>  1.0
 Test.instance =  2 =>  0.0

Instead of the integer (hard) predictions, I need probablistic predictions. There is an option -b for command line, but I couldn't find anything for direct usage of the function inside the code. Also, looked inside the code (https://github.com/bwaldvogel/liblinear-java/blob/master/src/main/java/de/bwaldvogel/liblinear/Predict.java); apparently there is no option for probabilistic prediction, via direct usage inside the code. Is that correct?

UPDATE: I ended up using the liblinear code form https://github.com/bwaldvogel/liblinear-java . In the file Predict.java I changed

private static boolean       flag_predict_probability = true;

to

private static boolean       flag_predict_probability = false;

and used

SolverType.L2R_LR

But still getting integer classes. Any idea?

Answer 1

To use probabilities one needs to change the code. The prediction is made inside the

public static double predictValues(Model model, Feature[] x, double[] dec_values) {

function inside Linear.java file:

    if (model.nr_class == 2) {
        System.out.println("Two classes "); 
        if (model.solverType.isSupportVectorRegression()) { 
            System.out.println("Support vector");
            return dec_values[0];
        }
        else { 
            System.out.println("Not Support vector");
            return (dec_values[0] > 0) ? model.label[0] : model.label[1];
        }

    } 

needs to be changed to

    if (model.nr_class == 2) {
        System.out.println("Two classes "); 
        if (model.solverType.isSupportVectorRegression()) { 
            System.out.println("Support vector");
            return dec_values[0];
        }
        else { 
            System.out.println("Not Support vector");
            return dec_values[0]; 
        }    
    } 

Note that the output is still not a probabilty, instead it is just a linear combination of weights and feature values. If you give it to softmax function it will become a probability in [0, 1].

Also, make sure to choose Logistic Regression:

     SolverType.L2R_LR