natas_hw
/
jidt
mirror of https://github.com/jlizier/jidt
0
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Correcting TE Kernel estimator's bias correction method to match Schreiber's approach to including contributions from other correlation sums when the full joint space has no neighbours.

This commit is contained in:
joseph.lizier 2014-08-05 02:04:56 +00:00
parent e03546c7b6
commit ab21895321
2 changed files with 43 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
java/source/infodynamics/measures/continuous/kernel/KernelEstimatorUniVariate.java
View File

@ -258,6 +258,10 @@ public class KernelEstimatorUniVariate {
System.out.println("Count after upper bin " + (bin + 1) + " = " + count);
}
// TODO Should we divide this result by the kernel width (in use)?
// Schreiber doesn't do this in the TE paper, but it is done in
// the Kaiser and Schreiber paper. It won't matter to TE, but
// will make a difference to individual entropy estimates.
return (double) count / (double) totalTimePointsCompared;
}

41
java/source/infodynamics/measures/continuous/kernel/TransferEntropyCalculatorKernel.java
View File

@ -237,6 +237,8 @@ public class TransferEntropyCalculatorKernel
* though the adverse correction of the bias correction is worse than the correction
* itself (because the probabilities being multiplied/divided are not independent),
* so recommend not to use this method.
* (Bias correction is implemented properly in the Kraskov et al.
* estimators, see {@link infodynamics.measures.continuous.kraskov.TransferEntropyCalculatorKraskov}
* </p>
* <p>It is implemented here for testing purposes only.</p>
*
@ -244,25 +246,60 @@ public class TransferEntropyCalculatorKernel
public double computeAverageLocalOfObservationsWithCorrection() throws Exception {
double te = 0.0;
int numNoNeighbours = 0;
double contributionsNoNeighbours = 0.0;
for (int b = 0; b < totalObservations; b++) {
TransferEntropyKernelCounts kernelCounts = teKernelEstimator.getCount(destPastVectors[b],
destNextValues[b], sourceValues[b], b);
double cont = 0.0;
if (kernelCounts.countNextPastSource > 0) {
// Original code:
/* if (kernelCounts.countNextPastSource > 0) {
cont = MathsUtils.digamma(kernelCounts.countNextPastSource) -
MathsUtils.digamma(kernelCounts.countPastSource) -
MathsUtils.digamma(kernelCounts.countNextPast) +
MathsUtils.digamma(kernelCounts.countPast);
} */
// But Schreiber confirmed to me that with dynamic correlation
// exclusion, when you may have no nearest neighbours,
// he was allowing contributions from other groups (e.g. countPastSource)
// to be added even if the full joint count was zero.
// Implement it like this:
// TODO Do we need to correct the totalObservations
// divisor for each digamma sum to account for this?
// (Schreiber does not do that; for the moment we'll accept that
// as the right approach)
if (kernelCounts.countPastSource > 0) {
cont -= MathsUtils.digamma(kernelCounts.countPastSource);
}
if (kernelCounts.countNextPast > 0) {
cont -= MathsUtils.digamma(kernelCounts.countNextPast);
}
if (kernelCounts.countPast > 0) {
cont += MathsUtils.digamma(kernelCounts.countPast);
}
if (kernelCounts.countNextPastSource > 0) {
cont += MathsUtils.digamma(kernelCounts.countNextPastSource);
} else {
// These contributions are from a set with no neighbours in
// the full joint space.
contributionsNoNeighbours += cont;
numNoNeighbours++;
}
te += cont;
/*
if (debug) {
System.out.println(b + ": " + logTerm + " -> " + (cont/Math.log(2.0)) + " -> sum: " + (te/Math.log(2.0)));
System.out.println(b + ": " + cont + " -> " + (cont/Math.log(2.0)) + " -> sum: " + (te/Math.log(2.0)));
}
*/
}
// Average it, and convert results to bytes
lastAverage = te / (double) totalObservations / Math.log(2.0);
if (debug) {
System.out.printf("TE=%.4f, with %d contributions from 0 neighbour sets being %.4f\n",
lastAverage,
numNoNeighbours,
contributionsNoNeighbours / (double) totalObservations / Math.log(2.0));
}
return lastAverage;
}