Skip to content

Commit

Permalink
Fix calculation of gap thresholds by correctly using actual value in …
Browse files Browse the repository at this point in the history
…ratio computations (#408)

In the calculation of `gapLow[y]` and `gapHigh[y]`, the expressions for the ratio-based thresholds were incorrectly using `Math.abs(a)` where `a = scale[y] * point[startPosition + y]`. Since `point[startPosition + y]` is the normalized value `(x - mean) / std`, multiplying by `scale[y]` (which is `std`) gives `(x - mean)`.

However, to accurately compute the thresholds based on the actual value `x`, we need to add back the mean (`shiftBase`). Therefore, `(a + shiftBase)` equals `(x - mean) + mean = x`.

The corrected code now uses `Math.abs(a + shiftBase)` in PredictorCorrector.

Testing done:
1. added a IT.

Signed-off-by: Kaituo Li <[email protected]>
  • Loading branch information
kaituo authored Oct 10, 2024
1 parent f2984b5 commit 35f4cf6
Show file tree
Hide file tree
Showing 12 changed files with 231 additions and 18 deletions.
10 changes: 5 additions & 5 deletions Java/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -157,7 +157,7 @@ vector data point, scores the data point, and then updates the model with this
point. The program output appends a column of anomaly scores to the input:

```text
$ java -cp core/target/randomcutforest-core-4.1.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner < ../example-data/rcf-paper.csv > example_output.csv
$ java -cp core/target/randomcutforest-core-4.2.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner < ../example-data/rcf-paper.csv > example_output.csv
$ tail example_output.csv
-5.0029,0.0170,-0.0057,0.8129401629464965
-4.9975,-0.0102,-0.0065,0.6591046054520615
Expand All @@ -176,8 +176,8 @@ read additional usage instructions, including options for setting model
hyperparameters, using the `--help` flag:

```text
$ java -cp core/target/randomcutforest-core-4.1.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner --help
Usage: java -cp target/random-cut-forest-4.1.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner [options] < input_file > output_file
$ java -cp core/target/randomcutforest-core-4.2.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner --help
Usage: java -cp target/random-cut-forest-4.2.0.jar com.amazon.randomcutforest.runner.AnomalyScoreRunner [options] < input_file > output_file
Compute scalar anomaly scores from the input rows and append them to the output rows.
Expand Down Expand Up @@ -239,14 +239,14 @@ framework. Build an executable jar containing the benchmark code by running
To invoke the full benchmark suite:

```text
% java -jar benchmark/target/randomcutforest-benchmark-4.1.0-jar-with-dependencies.jar
% java -jar benchmark/target/randomcutforest-benchmark-4.2.0-jar-with-dependencies.jar
```

The full benchmark suite takes a long time to run. You can also pass a regex at the command-line, then only matching
benchmark methods will be executed.

```text
% java -jar benchmark/target/randomcutforest-benchmark-4.1.0-jar-with-dependencies.jar RandomCutForestBenchmark\.updateAndGetAnomalyScore
% java -jar benchmark/target/randomcutforest-benchmark-4.2.0-jar-with-dependencies.jar RandomCutForestBenchmark\.updateAndGetAnomalyScore
```

[rcf-paper]: http://proceedings.mlr.press/v48/guha16.pdf
2 changes: 1 addition & 1 deletion Java/benchmark/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
<parent>
<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-benchmark</artifactId>
Expand Down
2 changes: 1 addition & 1 deletion Java/core/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
<parent>
<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-core</artifactId>
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -142,7 +142,7 @@ protected void updateTimestamps(long timestamp) {
* continuously since we are always counting missing values that should
* eventually be reset to zero. To address the issue, we add code in method
* updateForest to decrement numberOfImputed when we move to a new timestamp,
* provided there is no imputation. This ensures th e imputation fraction does
* provided there is no imputation. This ensures the imputation fraction does
* not increase as long as the imputation is continuing. This also ensures that
* the forest update decision, which relies on the imputation fraction,
* functions correctly. The forest is updated only when the imputation fraction
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
import static org.junit.jupiter.api.Assertions.assertDoesNotThrow;
import static org.junit.jupiter.api.Assertions.assertEquals;
import static org.junit.jupiter.api.Assertions.assertThrows;
import static org.junit.jupiter.api.Assertions.assertTrue;
import static org.mockito.ArgumentMatchers.any;
import static org.mockito.Mockito.mock;
import static org.mockito.Mockito.when;
Expand Down Expand Up @@ -343,7 +344,11 @@ public void ParallelTest(BiFunction<float[], float[], Double> distance) {
assertEquals(summary2.weightOfSamples, summary1.weightOfSamples, " sampling inconsistent");
assertEquals(summary2.summaryPoints.length, summary1.summaryPoints.length,
" incorrect length of typical points");
assertEquals(clusters.size(), summary1.summaryPoints.length);
// due to randomization, they might not equal
assertTrue(
Math.abs(clusters.size() - summary1.summaryPoints.length) <= 1,
"The difference between clusters.size() and summary1.summaryPoints.length should be at most 1"
);
double total = clusters.stream().map(ICluster::getWeight).reduce(0.0, Double::sum);
assertEquals(total, summary1.weightOfSamples, 1e-3);
// parallelization can produce reordering of merges
Expand Down
2 changes: 1 addition & 1 deletion Java/examples/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<parent>
<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-examples</artifactId>
Expand Down
2 changes: 1 addition & 1 deletion Java/parkservices/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
<parent>
<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-parkservices</artifactId>
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -464,19 +464,74 @@ protected <P extends AnomalyDescriptor> DiVector constructUncertaintyBox(float[]
double[] gapLow = new double[baseDimensions];
double[] gapHigh = new double[baseDimensions];
for (int y = 0; y < baseDimensions; y++) {
// 'a' represents the scaled value of the current point for dimension 'y'.
// Given that 'point[startPosition + y]' is the normalized value of the actual
// data point (x - mean) / std,
// and 'scale[y]' is the standard deviation (std), we have:
// a = std * ((x - mean) / std) = x - mean
double a = scale[y] * point[startPosition + y];

// 'shiftBase' is the shift value for dimension 'y', which is the mean (mean)
double shiftBase = shift[y];

// Initialize 'shiftAmount' to zero. This will account for numerical precision
// adjustments later
double shiftAmount = 0;

// If the mean ('shiftBase') is not zero, adjust 'shiftAmount' to account for
// numerical precision
if (shiftBase != 0) {
// 'shiftAmount' accounts for potential numerical errors due to shifting and
// scaling
shiftAmount += DEFAULT_NORMALIZATION_PRECISION * (scale[y] + Math.abs(shiftBase));
}

// Calculate the average L1 deviation along the path for dimension 'y'.
// This function computes the average absolute difference between successive
// values in the shingle,
// helping to capture recent fluctuations or trends in the data.
double pathGap = calculatePathDeviation(point, startPosition, y, baseDimension, differenced);

// 'noiseGap' is calculated based on the noise factor and the deviation for
// dimension 'y'.
// It represents the expected variation due to noise, scaled appropriately.
double noiseGap = noiseFactor * result.getDeviations()[baseDimension + y];

// 'gap' is the maximum of the scaled 'pathGap' and 'noiseGap', adjusted by
// 'shiftAmount'
// and a small constant to ensure it's not zero. This gap accounts for recent
// deviations and noise,
// and serves as a baseline threshold for detecting anomalies.
double gap = max(scale[y] * pathGap, noiseGap) + shiftAmount + DEFAULT_NORMALIZATION_PRECISION;
gapLow[y] = max(max(ignoreNearExpectedFromBelow[y], ignoreNearExpectedFromBelowByRatio[y] * Math.abs(a)),
gap);
gapHigh[y] = max(max(ignoreNearExpectedFromAbove[y], ignoreNearExpectedFromAboveByRatio[y] * Math.abs(a)),
gap);

// Compute 'gapLow[y]' and 'gapHigh[y]', which are thresholds to determine if
// the deviation is significant
// Since 'a = x - mean' and 'shiftBase = mean', then 'a + shiftBase = x - mean +
// mean = x'
// Therefore, 'Math.abs(a + shiftBase)' simplifies to the absolute value of the
// actual data point |x|
// For 'gapLow[y]', calculate the maximum of:
// - 'ignoreNearExpectedFromBelow[y]', an absolute threshold for ignoring small
// deviations below expected
// - 'ignoreNearExpectedFromBelowByRatio[y] * |x|', a relative threshold based
// on the actual value x
// - 'gap', the calculated deviation adjusted for noise and precision
// This ensures that minor deviations within the specified ratio or fixed
// threshold are ignored,
// reducing false positives.
gapLow[y] = max(max(ignoreNearExpectedFromBelow[y],
ignoreNearExpectedFromBelowByRatio[y] * (Math.abs(a + shiftBase))), gap);

// Similarly, for 'gapHigh[y]':
// - 'ignoreNearExpectedFromAbove[y]', an absolute threshold for ignoring small
// deviations above expected
// - 'ignoreNearExpectedFromAboveByRatio[y] * |x|', a relative threshold based
// on the actual value x
// - 'gap', the calculated deviation adjusted for noise and precision
// This threshold helps in ignoring anomalies that are within an acceptable
// deviation ratio from the expected value.
gapHigh[y] = max(max(ignoreNearExpectedFromAbove[y],
ignoreNearExpectedFromAboveByRatio[y] * (Math.abs(a + shiftBase))), gap);
}
return new DiVector(gapHigh, gapLow);
}
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,153 @@
/*
* Copyright 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License").
* You may not use this file except in compliance with the License.
* A copy of the License is located at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* or in the "license" file accompanying this file. This file is distributed
* on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either
* express or implied. See the License for the specific language governing
* permissions and limitations under the License.
*/

package com.amazon.randomcutforest.parkservices;

import static org.junit.jupiter.api.Assertions.assertTrue;

import java.time.LocalDateTime;
import java.time.temporal.ChronoUnit;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
import java.util.Random;
import java.util.Set;
import java.util.TreeSet;

import org.junit.jupiter.api.Test;

import com.amazon.randomcutforest.config.ForestMode;
import com.amazon.randomcutforest.config.Precision;
import com.amazon.randomcutforest.config.TransformMethod;

public class IgnoreTest {
@Test
public void testAnomalies() {
// Initialize the forest parameters
int shingleSize = 8;
int numberOfTrees = 50;
int sampleSize = 256;
Precision precision = Precision.FLOAT_32;
int baseDimensions = 1;

long count = 0;
int dimensions = baseDimensions * shingleSize;

// Build the ThresholdedRandomCutForest
ThresholdedRandomCutForest forest = new ThresholdedRandomCutForest.Builder<>().compact(true)
.dimensions(dimensions).randomSeed(0).numberOfTrees(numberOfTrees).shingleSize(shingleSize)
.sampleSize(sampleSize).precision(precision).anomalyRate(0.01).forestMode(ForestMode.STREAMING_IMPUTE)
.transformMethod(TransformMethod.NORMALIZE).autoAdjust(true)
.ignoreNearExpectedFromAboveByRatio(new double[] { 0.1 })
.ignoreNearExpectedFromBelowByRatio(new double[] { 0.1 }).build();

// Generate the list of doubles
List<Double> randomDoubles = generateUniformRandomDoubles();

// List to store detected anomaly indices
List<Integer> anomalies = new ArrayList<>();

// Process each data point through the forest
for (double val : randomDoubles) {
double[] point = new double[] { val };
long newStamp = 100 * count;

AnomalyDescriptor result = forest.process(point, newStamp);

if (result.getAnomalyGrade() != 0) {
anomalies.add((int) count);
}
++count;
}

// Expected anomalies
List<Integer> expectedAnomalies = Arrays.asList(273, 283, 505, 1323);

System.out.println("Anomalies detected at indices: " + anomalies);

// Verify that all expected anomalies are detected
assertTrue(anomalies.containsAll(expectedAnomalies),
"Anomalies detected do not contain all expected anomalies");
}

public static List<Double> generateUniformRandomDoubles() {
// Set fixed times for reproducibility
LocalDateTime startTime = LocalDateTime.of(2020, 1, 1, 0, 0, 0);
LocalDateTime endTime = LocalDateTime.of(2020, 1, 2, 0, 0, 0);
long totalIntervals = ChronoUnit.MINUTES.between(startTime, endTime);

// Generate timestamps (not used but kept for completeness)
List<LocalDateTime> timestamps = new ArrayList<>();
for (int i = 0; i < totalIntervals; i++) {
timestamps.add(startTime.plusMinutes(i));
}

// Initialize variables
Random random = new Random(0); // For reproducibility
double level = 0;
List<Double> logCounts = new ArrayList<>();

// Decide random change points where level will change
int numChanges = random.nextInt(6) + 5; // Random number between 5 and 10 inclusive

Set<Integer> changeIndicesSet = new TreeSet<>();
changeIndicesSet.add(0); // Ensure the first index is included

while (changeIndicesSet.size() < numChanges) {
int idx = random.nextInt((int) totalIntervals - 1) + 1; // Random index between 1 and totalIntervals -1
changeIndicesSet.add(idx);
}

List<Integer> changeIndices = new ArrayList<>(changeIndicesSet);

// Generate levels at each change point
List<Double> levels = new ArrayList<>();
for (int i = 0; i < changeIndices.size(); i++) {
if (i == 0) {
level = random.nextDouble() * 10; // Starting level between 0 and 10
} else {
double increment = -2 + random.nextDouble() * 7; // Random increment between -2 and 5
level = Math.max(0, level + increment);
}
levels.add(level);
}

// Now generate logCounts for each timestamp with even smoother transitions
int currentLevelIndex = 0;
for (int idx = 0; idx < totalIntervals; idx++) {
if (currentLevelIndex + 1 < changeIndices.size() && idx >= changeIndices.get(currentLevelIndex + 1)) {
currentLevelIndex++;
}
level = levels.get(currentLevelIndex);
double sineWave = Math.sin((idx % 300) * (Math.PI / 150)) * 0.05 * level;
double noise = (-0.01 * level) + random.nextDouble() * (0.02 * level); // Noise between -0.01*level and
// 0.01*level
double count = Math.max(0, level + sineWave + noise);
logCounts.add(count);
}

// Introduce controlled changes for anomaly detection testing
for (int changeIdx : changeIndices) {
if (changeIdx + 10 < totalIntervals) {
logCounts.set(changeIdx + 5, logCounts.get(changeIdx + 5) * 1.05); // 5% increase
logCounts.set(changeIdx + 10, logCounts.get(changeIdx + 10) * 1.10); // 10% increase
}
}

// Output the generated logCounts
System.out.println("Generated logCounts of size: " + logCounts.size());
return logCounts;
}
}
2 changes: 1 addition & 1 deletion Java/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@

<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
<packaging>pom</packaging>

<name>software.amazon.randomcutforest:randomcutforest</name>
Expand Down
2 changes: 1 addition & 1 deletion Java/serialization/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<parent>
<groupId>software.amazon.randomcutforest</groupId>
<artifactId>randomcutforest-parent</artifactId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-serialization</artifactId>
Expand Down
2 changes: 1 addition & 1 deletion Java/testutils/pom.xml
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
<parent>
<artifactId>randomcutforest-parent</artifactId>
<groupId>software.amazon.randomcutforest</groupId>
<version>4.1.0</version>
<version>4.2.0</version>
</parent>

<artifactId>randomcutforest-testutils</artifactId>
Expand Down

0 comments on commit 35f4cf6

Please sign in to comment.