edu.cmu.minorthird.classify.sequential
Interface ConfidenceReportingSequenceClassifier

All Superinterfaces:

SequenceClassifier

All Known Implementing Classes:

CMM

public interface ConfidenceReportingSequenceClassifier
extends SequenceClassifier

A sequence classifier that can report on the confidence associated with any subsequence of a sequential classification.

Author:

William Cohen

Method Summary

double

confidence(Instance[] sequence, ClassLabel[] predictedClasses, ClassLabel[] alternateClasses, int lo, int hi) Return a measure of confidence that the correct ClassLabels for indices lo...hi-1 of the 'sequence' are in fact the ones in 'predictedClasses[lo..hi-1]', rather than the ones given in in 'alternateClasses'.

Methods inherited from interface edu.cmu.minorthird.classify.sequential.SequenceClassifier

classification, explain, getExplanation

Method Detail

confidence

double confidence(Instance[] sequence,
                  ClassLabel[] predictedClasses,
                  ClassLabel[] alternateClasses,
                  int lo,
                  int hi)

Return a measure of confidence that the correct ClassLabels for indices lo...hi-1 of the 'sequence' are in fact the ones in 'predictedClasses[lo..hi-1]', rather than the ones given in in 'alternateClasses'.

You can think of 'alternateClasses' as a sort of template, indicating constraints on the sequence of predictions that are allowed. 'Confidence' is confidence of the predictedClasses relative to the best-scoring prediction that is consistent with the 'alternateClasses' template. NULL values in the alternateClasses array indicate positions for which no alternate ClassLabels are required. In this case, the confidence returned should be (approximately) the maximum over all possible values for the slots filled with NULL values.

For instance, if all alternateClasses values outside the range lo..hi-1 are NULL, then an HMM sequenceClassifier should return the ratio running forward-backward on the predictedClasses and alternateClasses, with only the values between lo..hi-1 fixed.

Since classifications in the sequence are not independent, the values of 'alternateClasses' in positions outside of lo..hi-1 may affect this confidence.

Parameters:

sequence - is a sequence of instances to classify

predictedClasses - is a classification of the sequence, as given by a call to 'classify'

alternateClasses - is another possible classification of the sequence.

lo - is an index into sequence[] (or equivalently, any of the other array arguments)

hi - is another index into sequence[]

Returns:

some value that monotonic in confidence