Examples with CBM edu.neu.ccs.pyramid.multilabel_classification.cbm.CBM used on opensource projects

Example 1 with CBM

use of edu.neu.ccs.pyramid.multilabel_classification.cbm.CBM in project pyramid by cheng-li.

the class KLDivergence method kl_conditional.

// empirical KL
public static double kl_conditional(MultiLabelClassifier.AssignmentProbEstimator multiLabelClassifier, MultiLabelClfDataSet dataSet) {
    Map<MultiLabel, Integer> q_z = new HashMap<MultiLabel, Integer>();
    Map<MultiLabel, HashMap<MultiLabel, Integer>> q_yz = new HashMap<MultiLabel, HashMap<MultiLabel, Integer>>();
    // get overall empirical distribution
    for (int i = 0; i < dataSet.getNumDataPoints(); ++i) {
        MultiLabel z = new MultiLabel(dataSet.getRow(i));
        MultiLabel y = dataSet.getMultiLabels()[i];
        if (q_z.containsKey(z)) {
            q_z.put(z, q_z.get(z) + 1);
        } else {
            q_z.put(z, 1);
        }
        if (!q_yz.containsKey(z)) {
            q_yz.put(z, new HashMap<MultiLabel, Integer>());
        }
        if (q_yz.get(z).containsKey(y)) {
            q_yz.get(z).put(y, q_yz.get(z).get(y) + 1);
        } else {
            q_yz.get(z).put(y, 1);
        }
    }
    // compute kl divergence
    double kl = 0.0;
    for (Map.Entry<MultiLabel, Integer> e1 : q_z.entrySet()) {
        double kl_y = 0.0;
        for (Map.Entry<MultiLabel, Integer> e2 : q_yz.get(e1.getKey()).entrySet()) {
            double empirical_prob_yz = (double) e2.getValue() / (double) e1.getValue();
            double log_estimated_prob_yz = multiLabelClassifier.predictLogAssignmentProb(e1.getKey().toVector(dataSet.getNumFeatures()), e2.getKey());
            kl_y += empirical_prob_yz * (Math.log(empirical_prob_yz) - log_estimated_prob_yz);
        }
        double empirical_prob_z = (double) e1.getValue() / (double) dataSet.getNumDataPoints();
        kl += empirical_prob_z * kl_y;
    }
    // Printing information if needed
    int occur_threshold = 10;
    double marginal_threshold = 0.01;
    for (Map.Entry<MultiLabel, Integer> e1 : q_z.entrySet()) {
        double[] marginals1 = new double[dataSet.getNumFeatures()];
        for (Map.Entry<MultiLabel, Integer> e2 : q_yz.get(e1.getKey()).entrySet()) {
            double estimated_prob_yz = multiLabelClassifier.predictAssignmentProb(e1.getKey().toVector(dataSet.getNumFeatures()), e2.getKey());
            double empirical_prob_yz = (double) e2.getValue() / (double) e1.getValue();
            if (e1.getValue() >= occur_threshold) {
                System.out.println("#z:" + e1.getValue() + ",z=" + e1.getKey().toStringWithExtLabels(dataSet.getLabelTranslator()) + "->{" + e2.getKey().toStringWithExtLabels(dataSet.getLabelTranslator()) + "},#y:" + e2.getValue() + ",p_y|z_empirical:" + empirical_prob_yz + ",p_y|z_estimated:" + estimated_prob_yz);
            }
            for (int i = 0; i < dataSet.getNumFeatures(); i++) {
                if (e2.getKey().matchClass(i)) {
                    marginals1[i] += e2.getValue();
                }
            }
        }
        if (e1.getValue() >= occur_threshold) {
            double estimated_prob_zz = multiLabelClassifier.predictAssignmentProb(e1.getKey().toVector(dataSet.getNumFeatures()), e1.getKey());
            System.out.println("p(y=z|z)=" + estimated_prob_zz);
            CBM cbm = (CBM) multiLabelClassifier;
            //                List<MultiLabel> sampled = cbm.samples(e1.getKey().toVector(dataSet.getNumFeatures()), 10);
            //                for (int i = 0; i < sampled.size(); ++i) {
            //                    double prob = multiLabelClassifier.predictAssignmentProb(e1.getKey().toVector(dataSet.getNumFeatures()), sampled.get(i));
            //                    System.out.println(sampled.get(i).toStringWithExtLabels(dataSet.getLabelTranslator()) + ":" + prob);
            //                }
            System.out.println("p_y|z_estimated marginals are: ");
            double[] marginals = cbm.predictClassProbs(e1.getKey().toVector(dataSet.getNumFeatures()));
            int[] order = ArgSort.argSortDescending(marginals);
            for (int i = 0; i < order.length; ++i) {
                if (marginals[order[i]] > marginal_threshold) {
                    System.out.println(dataSet.getLabelTranslator().toExtLabel(order[i]) + ":" + marginals[order[i]]);
                }
            }
            System.out.println("p_y|z_empirical marginals are: ");
            for (int i = 0; i < dataSet.getNumFeatures(); i++) {
                marginals1[i] /= (double) e1.getValue();
            }
            int[] order1 = ArgSort.argSortDescending(marginals1);
            for (int i = 0; i < order1.length; ++i) {
                if (marginals1[order1[i]] > marginal_threshold) {
                    System.out.println(dataSet.getLabelTranslator().toExtLabel(order1[i]) + ":" + marginals1[order1[i]]);
                }
            }
        }
    }
    System.out.println("LRs for each label:");
    CBM cbm = (CBM) multiLabelClassifier;
    Classifier.ProbabilityEstimator[] estimators = cbm.getBinaryClassifiers()[0];
    for (int i = 0; i < estimators.length; i++) {
        System.out.println("LR:" + dataSet.getLabelTranslator().toExtLabel(i));
        LogisticRegression lr = (LogisticRegression) estimators[i];
        Vector weight_vec = lr.getWeights().getWeightsWithoutBiasForClass(1);
        double[] weights = new double[weight_vec.size()];
        for (int j = 0; j < weight_vec.size(); j++) {
            weights[j] = weight_vec.get(j);
        }
        System.out.println("bias:" + lr.getWeights().getBiasForClass(1));
        int[] order2 = ArgSort.argSortDescending(weights);
        for (int j = 0; j < order2.length; ++j) {
            System.out.println(dataSet.getLabelTranslator().toExtLabel(order2[j]) + ":" + weights[order2[j]]);
        }
    }
    System.out.println("---");
    return kl;
}