Get raw percentages using API

[Peter Figliozzi]

I am currently using GNaiveBayes along with GAutoFilter. When I use my model for a prediction:

model.predict(query_vector, predicted_label);

Then the predicted_label holds the predicted class (0 or 1). Instead, I would like to get the numerical values, such p(0)=0.45 and p(1)=0.55, because I would like to set the classification threshold to another value. (I assume it uses 0.5 as default).

Is this possible with the existing Waffles classes, or do I need to make a new subclass of GNaiveBayes?

Thanks

Pete

[Mike Gashler]

The GNaiveBayes::predictDistribution method returns GPrediction objects. The GPrediction::asCategorical method returns categorical distributions. The, GCategoricalDistribution::likelihood method will tell you the likelihood of each category.

[Peter Figliozzi]

Thanks Mike.
Here's a code snippet for anyone else who is trying to figure this out:

void printRaw(double *query_vector, GAutoFilter & model){
  GPrediction prediction;
  model.predictDistribution(query_vector,  & prediction);
  GCategoricalDistribution * catDist = prediction.asCategorical();
  double p_bad = catDist->likelihood(0);
  double p_good = catDist->likelihood(1);
  cout << "p_good = " << p_good << "  p_bad = " << p_bad << endl;
}

This assumes you've created your model like so:

 // load the training data
  GMatrix training_matrix;
  training_matrix.loadArff("training.arff");
  // Split into separate data and class matrices, as required by Waffles algos
  // The "1" in the constructor means keep the last column as the label
  GDataColSplitter splitter(training_matrix, 1);
  GMatrix & features = splitter.features();
  GMatrix & labels = splitter.labels();
  // Create and train a model
  GAutoFilter model(new GNaiveBayes());
  model.train(features, labels);