Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ML] Avoid very low p-values if the term is only a tiny fraction of the foreground set #76764

Merged
merged 3 commits into from
Aug 20, 2021
Merged

[ML] Avoid very low p-values if the term is only a tiny fraction of the foreground set #76764

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a652d0a
Avoid very low p-values if category is only a tiny fraction of the fo…
tveasey Aug 20, 2021
d7d8fb3
Formatting
tveasey Aug 20, 2021
04f502c
Merge branch 'master' into sig-terms-p-value-for-low-fraction-categories
tveasey Aug 20, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 23 additions & 21 deletions x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/aggs/heuristic/PValueScore.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,48 +113,50 @@ public double getScore(long subsetFreq, long subsetSize, long supersetFreq, long
return 0.0;
}

// casting to `long` to round down to nearest whole number
double epsAllDocsInClass = (long)eps(allDocsInClass);
double epsAllDocsNotInClass = (long)eps(allDocsNotInClass);

docsContainTermInClass += epsAllDocsInClass;
docsContainTermNotInClass += epsAllDocsNotInClass;
allDocsInClass += epsAllDocsInClass;
allDocsNotInClass += epsAllDocsNotInClass;

// Adjust counts to ignore ratio changes which are less than 5%
// casting to `long` to round down to nearest whole number
docsContainTermNotInClass = (long)(Math.min(
1.05 * docsContainTermNotInClass,
docsContainTermInClass / allDocsInClass * allDocsNotInClass
) + 0.5);

// casting to `long` to round down to nearest whole number
double epsAllDocsInClass = (long)eps(allDocsInClass);
double epsAllDocsNotInClass = (long)eps(allDocsNotInClass);

if ((allDocsInClass + epsAllDocsInClass) > Long.MAX_VALUE
|| (docsContainTermInClass + epsAllDocsInClass) > Long.MAX_VALUE
|| (allDocsNotInClass + epsAllDocsNotInClass) > Long.MAX_VALUE
|| (docsContainTermNotInClass + epsAllDocsNotInClass) > Long.MAX_VALUE) {
if (allDocsInClass > Long.MAX_VALUE
|| docsContainTermInClass > Long.MAX_VALUE
|| allDocsNotInClass > Long.MAX_VALUE
|| docsContainTermNotInClass > Long.MAX_VALUE) {
throw new AggregationExecutionException(
"too many documents in background and foreground sets, further restrict sets for execution"
);
}

double v1 = new LongBinomialDistribution(
(long)(allDocsInClass + epsAllDocsInClass),
(docsContainTermInClass + epsAllDocsInClass)/(allDocsInClass + epsAllDocsInClass)
).logProbability((long)(docsContainTermInClass + epsAllDocsInClass));
(long)allDocsInClass, docsContainTermInClass / allDocsInClass
).logProbability((long)docsContainTermInClass);

double v2 = new LongBinomialDistribution(
(long)(allDocsNotInClass + epsAllDocsNotInClass),
(docsContainTermNotInClass + epsAllDocsNotInClass)/(allDocsNotInClass + epsAllDocsNotInClass)
).logProbability((long)(docsContainTermNotInClass + epsAllDocsNotInClass));
(long)allDocsNotInClass, docsContainTermNotInClass / allDocsNotInClass
).logProbability((long)docsContainTermNotInClass);

double p2 = (docsContainTermInClass + docsContainTermNotInClass + epsAllDocsNotInClass + epsAllDocsInClass)
/ (allDocsInClass + allDocsNotInClass + epsAllDocsNotInClass + epsAllDocsInClass);
double p2 = (docsContainTermInClass + docsContainTermNotInClass) / (allDocsInClass + allDocsNotInClass);

double v3 = new LongBinomialDistribution((long)(allDocsInClass + epsAllDocsInClass), p2)
.logProbability((long)(docsContainTermInClass + epsAllDocsInClass));
double v3 = new LongBinomialDistribution((long)allDocsInClass, p2)
.logProbability((long)docsContainTermInClass);

double v4 = new LongBinomialDistribution((long)(allDocsNotInClass + epsAllDocsNotInClass), p2)
.logProbability((long)(docsContainTermNotInClass + epsAllDocsNotInClass));
double v4 = new LongBinomialDistribution((long)allDocsNotInClass, p2)
.logProbability((long)docsContainTermNotInClass);

double logLikelihoodRatio = v1 + v2 - v3 - v4;
double pValue = CHI_SQUARED_DISTRIBUTION.survivalFunction(2.0 * logLikelihoodRatio);
return -FastMath.log(FastMath.max(pValue, Double.MIN_NORMAL));
return FastMath.max(-FastMath.log(FastMath.max(pValue, Double.MIN_NORMAL)), 0.0);
}

private double eps(double value) {
Expand Down
17 changes: 8 additions & 9 deletions ...k/plugin/ml/src/test/java/org/elasticsearch/xpack/ml/aggs/heuristic/PValueScoreTests.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,6 @@
import static org.hamcrest.Matchers.allOf;
import static org.hamcrest.Matchers.closeTo;
import static org.hamcrest.Matchers.equalTo;
import static org.hamcrest.Matchers.greaterThan;
import static org.hamcrest.Matchers.greaterThanOrEqualTo;
import static org.hamcrest.Matchers.lessThanOrEqualTo;

Expand Down Expand Up @@ -110,31 +109,31 @@ public void testPValueScore() {
);
assertThat(
FastMath.exp(-new PValueScore(false).getScore(10, 100, 10, 1000)),
closeTo(0.002988594884934073, eps)
closeTo(0.003972388976814195, eps)
);
assertThat(
FastMath.exp(-new PValueScore(false).getScore(10, 100, 200, 1000)),
closeTo(1.0, eps)
);
assertThat(
FastMath.exp(-new PValueScore(false).getScore(20, 10000, 5, 10000)),
closeTo(0.6309430298306147, eps)
closeTo(1.0, eps)
);
}

public void testSmallChanges() {
assertThat(
FastMath.exp(-new PValueScore(false).getScore(1, 4205, 0, 821496)),
closeTo(0.9572480202044421, eps)
closeTo(0.9999037287868853, eps)
);
// Same(ish) ratios
assertThat(
FastMath.exp(-new PValueScore(false).getScore(10, 4205, 195, 82149)),
closeTo(0.9893886454928338, eps)
closeTo(0.9995943820612134, eps)
);
assertThat(
FastMath.exp(-new PValueScore(false).getScore(10, 4205, 1950, 821496)),
closeTo(0.9867689169546193, eps)
closeTo(0.9999942565428899, eps)
);

// 4% vs 0%
Expand All @@ -145,12 +144,12 @@ public void testSmallChanges() {
// 4% vs 2%
assertThat(
FastMath.exp(-new PValueScore(false).getScore(168, 4205, 16429, 821496)),
closeTo(4.78464746423625e-06, eps)
closeTo(8.542608559219833e-5, eps)
);
// 4% vs 3.5%
assertThat(
FastMath.exp(-new PValueScore(false).getScore(168, 4205, 28752, 821496)),
closeTo(0.4728938449949742, eps)
closeTo(0.8833950526957098, eps)
);
}

Expand Down Expand Up @@ -186,7 +185,7 @@ public void testIncreasedSubsetIncreasedScore() {
for (int j = 1; j < 11; j++) {
double nextScore = getScore.apply(j*10L);
assertThat(nextScore, greaterThanOrEqualTo(0.0));
assertThat(nextScore, greaterThan(priorScore));
assertThat(nextScore, greaterThanOrEqualTo(priorScore));
priorScore = nextScore;
}
}
Expand Down