Click or drag to resize
Accord.NET (logo)

MultilabelProbabilityMethod Enumeration

Probability computation strategies for MultilabelSupportVectorMachine

Namespace:  Accord.MachineLearning.VectorMachines
Assembly:  Accord.MachineLearning (in Accord.MachineLearning.dll) Version: 3.7.0
public enum MultilabelProbabilityMethod
  Member nameValueDescription
PerClass0 Probabilities should be computed per-class, meaning the probabilities among all the classe should not need to sum up to one. This is the default when dealing with multi-label (as opposed to mult-class) classification problems.
SumsToOne1 Probabilities should be normalized to sum up to one. This will be done using the Softmax(Double) function considering the output probabilities of all classes.
SumsToOneWithEmphasisOnWinner2 Probabilities should be normalized to sum up to one. However, in a one-vs-rest setting, one of the machines will have been trained to specifically distinguish between the winning class and the rest of the classes. As such the output of this class will be used to determine the true probability of the winning class, and the complement of this probabilitiy will be divided among the restant of the losing classses. The probabilities of the losing classes will be determined using a softmax.
See Also