def testSkipRecords(self):
"""
This calls estimateAnomalyLikelihoods with various values of skipRecords
"""
# Check happy path
data1 = _generateSampleData(mean=0.1)[0:200]
data1 = data1 + (_generateSampleData(mean=0.9)[0:200])
likelihoods, _, estimatorParams = (an.estimateAnomalyLikelihoods(
data1, skipRecords=200))
# Check results are correct, i.e. we are actually skipping the first 50
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], 0.9, epsilon=0.1)
# Check case where skipRecords > num records
# In this case a null distribution should be returned which makes all
# the likelihoods reasonably high
likelihoods, _, estimatorParams = (an.estimateAnomalyLikelihoods(
data1, skipRecords=500))
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.3 * len(likelihoods))
# Check the case where skipRecords == num records
likelihoods, _, estimatorParams = (an.estimateAnomalyLikelihoods(
data1, skipRecords=len(data1)))
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.3 * len(likelihoods))
def testSkipRecords(self):
"""
This calls estimateAnomalyLikelihoods with various values of skipRecords
"""
# Check happy path
data1 = _generateSampleData(mean=0.1)[0:200]
data1 = data1 + (_generateSampleData(mean=0.9)[0:200])
likelihoods, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data1, skipRecords=200)
)
# Check results are correct, i.e. we are actually skipping the first 50
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], 0.9, epsilon=0.1)
# Check case where skipRecords > num records
# In this case a null distribution should be returned which makes all
# the likelihoods reasonably high
likelihoods, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data1, skipRecords=500)
)
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.3 * len(likelihoods))
# Check the case where skipRecords == num records
likelihoods, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data1, skipRecords=len(data1))
)
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.3 * len(likelihoods))
def testVeryFewScores(self):
"""
This calls estimateAnomalyLikelihoods and updateAnomalyLikelihoods
with one or no scores.
"""
# Generate an estimate using two data points
data1 = _generateSampleData(mean=42.0, variance=1e-10)
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(data1[0:2]))
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the estimated mean is that value
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], data1[0][2])
# Can't generate an estimate using no data points
data1 = numpy.zeros(0)
with self.assertRaises(ValueError):
an.estimateAnomalyLikelihoods(data1)
# Can't update with no scores
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, estimatorParams)
def testVeryFewScores(self):
"""
This calls estimateAnomalyLikelihoods and updateAnomalyLikelihoods
with one or no scores.
"""
# Generate an estimate using two data points
data1 = _generateSampleData(mean=42.0, variance=1e-10)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:2])
)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the estimated mean is that value
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], data1[0][2])
# Can't generate an estimate using no data points
data1 = numpy.zeros(0)
with self.assertRaises(ValueError):
an.estimateAnomalyLikelihoods(data1)
# Can't update with no scores
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, estimatorParams)
def testFlatAnomalyScores(self):
"""
This calls estimateAnomalyLikelihoods with flat distributions and
ensures things don't crash.
"""
# Generate an estimate using fake distribution of anomaly scores.
data1 = _generateSampleData(mean=42.0, variance=1e-10)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1000])
)
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
## Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], data1[0][2])
# If you deviate from the mean, you should get probability 0
# Test this by sending in just slightly different values.
data2 = _generateSampleData(mean=42.5, variance=1e-10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data2[0:10], estimatorParams)
)
# The likelihoods should go to zero very quickly
self.assertLessEqual(likelihoods2.sum(), 0.01)
# Test edge case where anomaly scores are very close to 0
# In this case we don't let likelihood to get too low. An average
# anomaly score of 0.1 should be essentially zero, but an average
# of 0.04 should be higher
data3 = _generateSampleData(mean=0.01, variance=1e-6)
_, _, estimatorParams3 = (
an.estimateAnomalyLikelihoods(data3[0:1000])
)
data4 = _generateSampleData(mean=0.1, variance=1e-6)
likelihoods4, _, estimatorParams4 = (
an.updateAnomalyLikelihoods(data4[0:20], estimatorParams3)
)
# Average of 0.1 should go to zero
self.assertLessEqual(likelihoods4[10:].mean(), 0.002)
data5 = _generateSampleData(mean=0.05, variance=1e-6)
likelihoods5, _, _ = (
an.updateAnomalyLikelihoods(data5[0:20], estimatorParams4)
)
# The likelihoods should be low but not near zero
self.assertLessEqual(likelihoods5[10:].mean(), 0.28)
self.assertGreater(likelihoods5[10:].mean(), 0.015)
def testFlatAnomalyScores(self):
"""
This calls estimateAnomalyLikelihoods with flat distributions and
ensures things don't crash.
"""
# Generate an estimate using fake distribution of anomaly scores.
data1 = _generateSampleData(mean=42.0, variance=1e-10)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1000])
)
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
## Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"], data1[0][2])
# If you deviate from the mean, you should get probability 0
# Test this by sending in just slightly different values.
data2 = _generateSampleData(mean=42.5, variance=1e-10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data2[0:10], estimatorParams)
)
# The likelihoods should go to zero very quickly
self.assertLessEqual(likelihoods2.sum(), 0.01)
# Test edge case where anomaly scores are very close to 0
# In this case we don't let likelihood to get too low. An average
# anomaly score of 0.1 should be essentially zero, but an average
# of 0.04 should be higher
data3 = _generateSampleData(mean=0.01, variance=1e-6)
_, _, estimatorParams3 = (
an.estimateAnomalyLikelihoods(data3[0:1000])
)
data4 = _generateSampleData(mean=0.1, variance=1e-6)
likelihoods4, _, estimatorParams4 = (
an.updateAnomalyLikelihoods(data4[0:20], estimatorParams3)
)
# Average of 0.1 should go to zero
self.assertLessEqual(likelihoods4[10:].mean(), 0.002)
data5 = _generateSampleData(mean=0.05, variance=1e-6)
likelihoods5, _, _ = (
an.updateAnomalyLikelihoods(data5[0:20], estimatorParams4)
)
# The likelihoods should be low but not near zero
self.assertLessEqual(likelihoods5[10:].mean(), 0.28)
self.assertGreater(likelihoods5[10:].mean(), 0.015)
def likelihood(self, value, anomalyScore, dttm):
"""
Given the current metric value, plus the current anomaly score, output the
anomalyLikelihood for this record.
"""
dataPoint = (dttm, value, anomalyScore)
# We ignore the first probationaryPeriod data points
if len(self._historicalScores) < self._probationaryPeriod:
likelihood = 0.5
else:
# On a rolling basis we re-estimate the distribution every 100 iterations
if self._distribution is None or (self._iteration % 100 == 0):
_, _, self._distribution = (
anomaly_likelihood.estimateAnomalyLikelihoods(
self._historicalScores,
skipRecords = self._numentaLearningPeriod)
)
likelihoods, _, self._distribution = (
anomaly_likelihood.updateAnomalyLikelihoods([dataPoint],
self._distribution)
)
likelihood = 1.0 - likelihoods[0]
# Before we exit update historical scores and iteration
self._historicalScores.append(dataPoint)
self._iteration += 1
return likelihood
def testCaseIncreasedAnomalyScore(self):
"""
Test F: small anomaly score every 20 records, but then a large one when you
would expect a small one. This should be anomalous.
"""
# Generate initial data
data = []
data = self._addSampleData(data, spikePeriod=20,
spikeValue=0.4, numSamples=1000)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data)
)
# Now feed in a more frequent distribution
data = self._addSampleData(spikePeriod=20, spikeValue=1.0,
numSamples=100)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# We should detect highly unusual behavior
self.assertTrue(likelihoods2.min() < 0.0003)
# We should detect it pretty often
self.assertTrue((likelihoods2 < 0.0003).sum() > 40)
def testCaseIncreasedSpikeFrequency(self):
"""
Test E: bunches of anomalies every 20 records that become even more
frequent. This should be anomalous.
"""
# Generate initial data
data = []
for _ in range(30):
data = self._addSampleData(data, spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# Now feed in a more frequent distribution
data = self._addSampleData(spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=1, numSamples=10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The likelihood should become anomalous but only near the end
self.assertTrue(likelihoods2[0:30].min() > 0.01)
self.assertTrue(likelihoods2[-5:].min() < 0.002)
def testCaseContinuousBunchesOfSpikes(self):
"""
Test D: bunches of anomalies every 20 records that continue. This should not
be anomalous.
"""
# Generate initial data
data = []
for _ in range(30):
data = self._addSampleData(data, spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# Now feed in the same distribution
data = self._addSampleData(spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The likelihood should be reasonable high everywhere
self.assertTrue(likelihoods2.min() > 0.01)
def testCaseUnusuallyHighSpikeFrequency(self):
"""
Test B: one anomaly spike every 20 records. Then we suddenly get a bunch
in a row. The likelihood of those spikes should be low.
"""
data = self._addSampleData(spikePeriod=20, numSamples=1019)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# If we continue to see the same distribution, we should get reasonable
# likelihoods
data = self._addSampleData(numSamples=119, spikePeriod=20)
likelihoods1, _, estimatorParams1 = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The minimum likelihood should be reasonably high
self.assertTrue((likelihoods1.min() > 0.1 ))
data = self._addSampleData(numSamples=20, spikePeriod=2)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams1)
)
# The likelihood once you get past the initial averaging should be very low.
self.assertTrue((likelihoods2[5:].sum() / 15.0) < 0.001)
def testCaseIncreasedSpikeFrequency(self):
"""
Test E: bunches of anomalies every 20 records that become even more
frequent. This should be anomalous.
"""
# Generate initial data
data = []
for _ in range(30):
data = self._addSampleData(data, spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# Now feed in a more frequent distribution
data = self._addSampleData(spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=1, numSamples=10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The likelihood should become anomalous but only near the end
self.assertTrue(likelihoods2[0:30].min() > 0.01)
self.assertTrue(likelihoods2[-5:].min() < 0.002)
def testEstimateAnomalyLikelihoods(self):
"""
This calls estimateAnomalyLikelihoods to estimate the distribution on fake
data and validates the results
"""
# Generate an estimate using fake distribution of anomaly scores.
data1 = _generateSampleData(mean=0.2)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1000]))
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the sum is correct
avgParams = estimatorParams["movingAverage"]
total = 0
for v in avgRecordList:
total = total + v[2]
self.assertTrue(avgParams["total"], total)
# Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"],
total / float(len(avgRecordList)))
# Number of points with lower than 2% probability should be pretty low
# but not zero. Can't use exact 2% here due to random variations
self.assertLessEqual(numpy.sum(likelihoods < 0.02), 50)
self.assertGreaterEqual(numpy.sum(likelihoods < 0.02), 1)
def testUpdateAnomalyLikelihoods(self):
"""
A slight more complex test. This calls estimateAnomalyLikelihoods
to estimate the distribution on fake data, followed by several calls
to updateAnomalyLikelihoods.
"""
# ------------------------------------------
# Step 1. Generate an initial estimate using fake distribution of anomaly
# scores.
data1 = _generateSampleData(mean=0.2)[0:1000]
_, _, estimatorParams = an.estimateAnomalyLikelihoods(data1, averagingWindow=5)
# ------------------------------------------
# Step 2. Generate some new data with a higher average anomaly
# score. Using the estimator from step 1, to compute likelihoods. Now we
# should see a lot more anomalies.
data2 = _generateSampleData(mean=0.6)[0:300]
likelihoods2, avgRecordList2, estimatorParams2 = an.updateAnomalyLikelihoods(data2, estimatorParams)
self.assertEqual(len(likelihoods2), len(data2))
self.assertEqual(len(avgRecordList2), len(data2))
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# The new running total should be different
self.assertNotEqual(estimatorParams2["movingAverage"]["total"], estimatorParams["movingAverage"]["total"])
# We should have many more samples where likelihood is < 0.01, but not all
self.assertGreaterEqual(numpy.sum(likelihoods2 < 0.01), 25)
self.assertLessEqual(numpy.sum(likelihoods2 < 0.01), 250)
# ------------------------------------------
# Step 3. Generate some new data with the expected average anomaly score. We
# should see fewer anomalies than in Step 2.
data3 = _generateSampleData(mean=0.2)[0:1000]
likelihoods3, avgRecordList3, estimatorParams3 = an.updateAnomalyLikelihoods(data3, estimatorParams2)
self.assertEqual(len(likelihoods3), len(data3))
self.assertEqual(len(avgRecordList3), len(data3))
self.assertTrue(an.isValidEstimatorParams(estimatorParams3))
# The new running total should be different
self.assertNotEqual(estimatorParams3["movingAverage"]["total"], estimatorParams["movingAverage"]["total"])
self.assertNotEqual(estimatorParams3["movingAverage"]["total"], estimatorParams2["movingAverage"]["total"])
# We should have a small number samples where likelihood is < 0.02, but at
# least one
self.assertGreaterEqual(numpy.sum(likelihoods3 < 0.01), 1)
self.assertLessEqual(numpy.sum(likelihoods3 < 0.01), 100)
# ------------------------------------------
# Step 4. Validate that sending data incrementally is the same as sending
# in one batch
allData = data1
allData.extend(data2)
allData.extend(data3)
# Compute moving average of all the data and check it's the same
_, historicalValuesAll, totalAll = an._anomalyScoreMovingAverage(allData, windowSize=5)
self.assertEqual(sum(historicalValuesAll), sum(estimatorParams3["movingAverage"]["historicalValues"]))
self.assertEqual(totalAll, estimatorParams3["movingAverage"]["total"])
def likelihood(self, value, anomalyScore, dttm):
"""
Given the current metric value, plus the current anomaly score, output the
anomalyLikelihood for this record.
"""
dataPoint = (dttm, value, anomalyScore)
# We ignore the first probationaryPeriod data points
if len(self._historicalScores) < self._probationaryPeriod:
likelihood = 0.5
else:
# On a rolling basis we re-estimate the distribution every 100 iterations
if self._distribution is None or (self._iteration % 100 == 0):
_, _, self._distribution = (
anomaly_likelihood.estimateAnomalyLikelihoods(
self._historicalScores,
skipRecords = self._numentaLearningPeriod)
)
likelihoods, _, self._distribution = (
anomaly_likelihood.updateAnomalyLikelihoods([dataPoint],
self._distribution)
)
likelihood = 1.0 - likelihoods[0]
# Before we exit update historical scores and iteration
self._historicalScores.append(dataPoint)
self._iteration += 1
return likelihood
def testEstimateAnomalyLikelihoodsMalformedRecords(self):
"""
This calls estimateAnomalyLikelihoods with malformed records, which should
be quietly skipped.
"""
# Generate an estimate using fake distribution of anomaly scores.
data1 = _generateSampleData(mean=0.2)
data1 = data1 + [(2, 2), (2, 2, 2, 2), (), (2)] # Malformed records
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1000])
)
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the sum is correct
avgParams = estimatorParams["movingAverage"]
total = 0
for v in avgRecordList:
total = total + v[2]
self.assertTrue(avgParams["total"], total)
# Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"],
total / float(len(avgRecordList)))
def testCaseUnusuallyHighSpikeFrequency(self):
"""
Test B: one anomaly spike every 20 records. Then we suddenly get a bunch
in a row. The likelihood of those spikes should be low.
"""
data = self._addSampleData(spikePeriod=20, numSamples=1019)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# If we continue to see the same distribution, we should get reasonable
# likelihoods
data = self._addSampleData(numSamples=119, spikePeriod=20)
likelihoods1, _, estimatorParams1 = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The minimum likelihood should be reasonably high
self.assertTrue((likelihoods1.min() > 0.1 ))
data = self._addSampleData(numSamples=20, spikePeriod=2)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams1)
)
# The likelihood once you get past the initial averaging should be very low.
self.assertTrue((likelihoods2[5:].sum() / 15.0) < 0.001)
def testCaseIncreasedAnomalyScore(self):
"""
Test F: small anomaly score every 20 records, but then a large one when you
would expect a small one. This should be anomalous.
"""
# Generate initial data
data = []
data = self._addSampleData(data,
spikePeriod=20,
spikeValue=0.4,
numSamples=1000)
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(data))
# Now feed in a more frequent distribution
data = self._addSampleData(spikePeriod=20,
spikeValue=1.0,
numSamples=100)
likelihoods2, _, _ = (an.updateAnomalyLikelihoods(
data, estimatorParams))
# We should detect highly unusual behavior
self.assertTrue(likelihoods2.min() < 0.0003)
# We should detect it pretty often
self.assertTrue((likelihoods2 < 0.0003).sum() > 40)
def testEstimateAnomalyLikelihoods(self):
"""
This calls estimateAnomalyLikelihoods to estimate the distribution on fake
data and validates the results
"""
# Generate an estimate using fake distribution of anomaly scores.
data1 = _generateSampleData(mean=0.2)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1000])
)
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the sum is correct
avgParams = estimatorParams["movingAverage"]
total = 0
for v in avgRecordList:
total = total + v[2]
self.assertTrue(avgParams["total"], total)
# Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"],
total / float(len(avgRecordList)))
# Number of points with lower than 2% probability should be pretty low
# but not zero. Can't use exact 2% here due to random variations
self.assertLessEqual(numpy.sum(likelihoods < 0.02), 50)
self.assertGreaterEqual(numpy.sum(likelihoods < 0.02), 1)
def testCaseContinuousBunchesOfSpikes(self):
"""
Test D: bunches of anomalies every 20 records that continue. This should not
be anomalous.
"""
# Generate initial data
data = []
for _ in range(30):
data = self._addSampleData(data, spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# Now feed in the same distribution
data = self._addSampleData(spikePeriod=0, numSamples=30)
data = self._addSampleData(data, spikePeriod=3, numSamples=10)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The likelihood should be reasonable high everywhere
self.assertTrue(likelihoods2.min() > 0.01)
def testEstimateAnomalyLikelihoodsMalformedRecords(self):
"""
This calls estimateAnomalyLikelihoods with malformed records, which should
be quietly skipped.
"""
# Generate a fake distribution of anomaly scores, and add malformed records
data1 = _generateSampleData(mean=0.2)
data1 = data1[0:1000] + [(2, 2)] + [(2, 2, 2, 2)] + [()] + [(2)]
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data1[0:1004]))
self.assertEqual(len(likelihoods), 1000)
self.assertEqual(len(avgRecordList), 1000)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Check that the sum is correct
avgParams = estimatorParams["movingAverage"]
total = 0
for v in avgRecordList:
total = total + v[2]
self.assertTrue(avgParams["total"], total)
# Check that the estimated mean is correct
dParams = estimatorParams["distribution"]
self.assertWithinEpsilon(dParams["mean"],
total / float(len(avgRecordList)))
def testEstimateAnomalyLikelihoodsCategoryValues(self):
start = datetime.datetime(2017, 1, 1, 0, 0, 0)
delta = datetime.timedelta(minutes=5)
dts = [start + (i * delta) for i in xrange(10)]
values = ["a", "b", "c", "d", "e"] * 2
rawScores = [0.1 * i for i in xrange(10)]
data = zip(dts, values, rawScores)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data))
self.assertEqual(len(likelihoods), 10)
self.assertEqual(len(avgRecordList), 10)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
def testCaseSingleSpike(self):
"""
No anomalies, and then you see a single spike. The likelihood of that
spike should be 0
"""
data = self._addSampleData(spikePeriod=0, numSamples=1000)
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(data[0:1000]))
data = self._addSampleData(numSamples=1, spikePeriod=1)
likelihoods1, _, _ = (an.updateAnomalyLikelihoods(
data, estimatorParams))
self.assertWithinEpsilon(likelihoods1[0], 0.0)
def testEstimateAnomalyLikelihoodsCategoryValues(self):
start = datetime.datetime(2017, 1, 1, 0, 0, 0)
delta = datetime.timedelta(minutes=5)
dts = [start + (i * delta) for i in xrange(10)]
values = ["a", "b", "c", "d", "e"] * 2
rawScores = [0.1 * i for i in xrange(10)]
data = zip(dts, values, rawScores)
likelihoods, avgRecordList, estimatorParams = (
an.estimateAnomalyLikelihoods(data)
)
self.assertEqual(len(likelihoods), 10)
self.assertEqual(len(avgRecordList), 10)
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
def testCaseMissingSpike(self):
"""
Test C: one anomaly every 20 records, but then see none. The likelihood
at the end should be very low.
"""
# Initial data
data = self._addSampleData(spikePeriod=20, numSamples=1019)
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(data[0:1000]))
# Now feed in none
data = self._addSampleData(numSamples=100, spikePeriod=0)
likelihoods2, _, _ = (an.updateAnomalyLikelihoods(
data, estimatorParams))
# The likelihood once you get past the initial averaging should be very low.
self.assertTrue((likelihoods2[5:].sum() / 15.0) < 0.0001)
def testCaseSingleSpike(self):
"""
No anomalies, and then you see a single spike. The likelihood of that
spike should be 0
"""
data = self._addSampleData(spikePeriod=0, numSamples=1000)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
data = self._addSampleData(numSamples=1, spikePeriod=1)
likelihoods1, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
self.assertWithinEpsilon(likelihoods1[0], 0.0)
def testFlatMetricScores(self):
"""
This calls estimateAnomalyLikelihoods with flat metric values. In this case
we should use the null distribution, which gets reasonably high likelihood
for everything.
"""
# Generate samples with very flat metric values
data1 = _generateSampleData(metricMean=42.0, metricVariance=1e-10)[0:1000]
likelihoods, _, estimatorParams = an.estimateAnomalyLikelihoods(data1)
# Check that we do indeed get reasonable likelihood values
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.4 * len(likelihoods))
# Check that we do indeed get null distribution
self.assertDictEqual(estimatorParams["distribution"], an.nullDistribution())
def testBadParams(self):
"""
Calls updateAnomalyLikelihoods with bad params.
"""
# Generate an estimate using one data point
data1 = _generateSampleData(mean=42.0, variance=1e-10)
_, _, estimatorParams = an.estimateAnomalyLikelihoods(data1[0:1])
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Can't pass in a bad params structure
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, {"haha": "heehee"})
# Can't pass in something not a dict
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, 42.0)
def testBadParams(self):
"""
Calls updateAnomalyLikelihoods with bad params.
"""
# Generate an estimate using one data point
data1 = _generateSampleData(mean=42.0, variance=1e-10)
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(data1[0:1]))
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# Can't pass in a bad params structure
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, {"haha": "heehee"})
# Can't pass in something not a dict
with self.assertRaises(ValueError):
an.updateAnomalyLikelihoods(data1, 42.0)
def testCaseMissingSpike(self):
"""
Test C: one anomaly every 20 records, but then see none. The likelihood
at the end should be very low.
"""
# Initial data
data = self._addSampleData(spikePeriod=20, numSamples=1019)
_, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data[0:1000])
)
# Now feed in none
data = self._addSampleData(numSamples=100, spikePeriod=0)
likelihoods2, _, _ = (
an.updateAnomalyLikelihoods(data, estimatorParams)
)
# The likelihood once you get past the initial averaging should be very low.
self.assertTrue((likelihoods2[5:].sum() / 15.0) < 0.0001)
def testFlatMetricScores(self):
"""
This calls estimateAnomalyLikelihoods with flat metric values. In this case
we should use the null distribution, which gets reasonably high likelihood
for everything.
"""
# Generate samples with very flat metric values
data1 = _generateSampleData(metricMean=42.0,
metricVariance=1e-10)[0:1000]
likelihoods, _, estimatorParams = (
an.estimateAnomalyLikelihoods(data1))
# Check that we do indeed get reasonable likelihood values
self.assertEqual(len(likelihoods), len(data1))
self.assertTrue(likelihoods.sum() >= 0.4 * len(likelihoods))
# Check that we do indeed get null distribution
self.assertDictEqual(estimatorParams["distribution"],
an.nullDistribution())
def testNABAnomalyLikelihood(self):
"""
Tests the specific calls to nupic/algorithms/anomaly_likelihood as they"re
made in "NAB/detectors/numenta/numenta_detector.py".
Note "NAB/.../numenta_detector.py" has its own class AnomalyLikelihood,
different from nupic/algorithms/anomaly_likelihood.AnomalyLikelihood, but
which calls the functions estimateAnomalyLikelihoods() and
updateAnomalyLikelihoods() from "nupic/algorithms/anomaly_likelihood.py".
"""
# AnomalyLikelihood object initial values
iteration = 0
probationaryPeriod = 4
historicalScores = []
likelihoodList = []
for dataPoint in self.data:
# Ignore the first probationaryPeriod data points
if len(historicalScores) < probationaryPeriod:
likelihood = 0.5
else:
if iteration % 4 == 0:
_, _, distribution = an.estimateAnomalyLikelihoods(
historicalScores, skipRecords=probationaryPeriod)
likelihoods, _, distribution = an.updateAnomalyLikelihoods(
[dataPoint], distribution)
likelihood = 1.0 - likelihoods[0]
historicalScores.append(dataPoint)
iteration += 1
likelihoodList.append(likelihood)
truthLikelihoodList = [
0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.044565462999999972,
0.044565462999999972, 0.044565462999999972, 0.044565462999999972,
0.90319951499999995, 0.90319951499999995, 0.90319951499999995,
0.90319951499999995, 0.78814460099999994, 0.78814460099999994,
0.78814460099999994, 0.78814460099999994
]
for i in xrange(len(likelihoodList)):
self.assertAlmostEqual(likelihoodList[i],
truthLikelihoodList[i],
msg="unequal values are at index %i" % i)
def testNABAnomalyLikelihood(self):
"""
Tests the specific calls to nupic/algorithms/anomaly_likelihood as they"re
made in "NAB/detectors/numenta/numenta_detector.py".
Note "NAB/.../numenta_detector.py" has its own class AnomalyLikelihood,
different from nupic/algorithms/anomaly_likelihood.AnomalyLikelihood, but
which calls the functions estimateAnomalyLikelihoods() and
updateAnomalyLikelihoods() from "nupic/algorithms/anomaly_likelihood.py".
"""
# AnomalyLikelihood object initial values
iteration = 0
probationaryPeriod = 4
historicalScores = []
likelihoodList = []
for dataPoint in self.data:
# Ignore the first probationaryPeriod data points
if len(historicalScores) < probationaryPeriod:
likelihood = 0.5
else:
if iteration % 4 == 0:
_, _, distribution = an.estimateAnomalyLikelihoods(
historicalScores,
skipRecords = probationaryPeriod)
likelihoods, _, distribution = an.updateAnomalyLikelihoods(
[dataPoint], distribution)
likelihood = 1.0 - likelihoods[0]
historicalScores.append(dataPoint)
iteration += 1
likelihoodList.append(likelihood)
truthLikelihoodList = [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5,
0.044565462999999972, 0.044565462999999972,
0.044565462999999972, 0.044565462999999972,
0.90319951499999995, 0.90319951499999995,
0.90319951499999995, 0.90319951499999995,
0.78814460099999994, 0.78814460099999994,
0.78814460099999994, 0.78814460099999994]
for i in xrange(len(likelihoodList)):
self.assertAlmostEqual(likelihoodList[i], truthLikelihoodList[i],
msg="unequal values are at index %i" % i)
def testUpdateAnomalyLikelihoods(self):
"""
A slight more complex test. This calls estimateAnomalyLikelihoods
to estimate the distribution on fake data, followed by several calls
to updateAnomalyLikelihoods.
"""
#------------------------------------------
# Step 1. Generate an initial estimate using fake distribution of anomaly
# scores.
data1 = _generateSampleData(mean=0.2)[0:1000]
_, _, estimatorParams = (an.estimateAnomalyLikelihoods(
data1, averagingWindow=5))
#------------------------------------------
# Step 2. Generate some new data with a higher average anomaly
# score. Using the estimator from step 1, to compute likelihoods. Now we
# should see a lot more anomalies.
data2 = _generateSampleData(mean=0.6)[0:300]
likelihoods2, avgRecordList2, estimatorParams2 = (
an.updateAnomalyLikelihoods(data2, estimatorParams))
self.assertEqual(len(likelihoods2), len(data2))
self.assertEqual(len(avgRecordList2), len(data2))
self.assertTrue(an.isValidEstimatorParams(estimatorParams))
# The new running total should be different
self.assertNotEqual(estimatorParams2["movingAverage"]["total"],
estimatorParams["movingAverage"]["total"])
# We should have many more samples where likelihood is < 0.01, but not all
self.assertGreaterEqual(numpy.sum(likelihoods2 < 0.01), 25)
self.assertLessEqual(numpy.sum(likelihoods2 < 0.01), 250)
#------------------------------------------
# Step 3. Generate some new data with the expected average anomaly score. We
# should see fewer anomalies than in Step 2.
data3 = _generateSampleData(mean=0.2)[0:1000]
likelihoods3, avgRecordList3, estimatorParams3 = (
an.updateAnomalyLikelihoods(data3, estimatorParams2))
self.assertEqual(len(likelihoods3), len(data3))
self.assertEqual(len(avgRecordList3), len(data3))
self.assertTrue(an.isValidEstimatorParams(estimatorParams3))
# The new running total should be different
self.assertNotEqual(estimatorParams3["movingAverage"]["total"],
estimatorParams["movingAverage"]["total"])
self.assertNotEqual(estimatorParams3["movingAverage"]["total"],
estimatorParams2["movingAverage"]["total"])
# We should have a small number samples where likelihood is < 0.02, but at
# least one
self.assertGreaterEqual(numpy.sum(likelihoods3 < 0.01), 1)
self.assertLessEqual(numpy.sum(likelihoods3 < 0.01), 100)
#------------------------------------------
# Step 4. Validate that sending data incrementally is the same as sending
# in one batch
allData = data1
allData.extend(data2)
allData.extend(data3)
# Compute moving average of all the data and check it's the same
_, historicalValuesAll, totalAll = (an._anomalyScoreMovingAverage(
allData, windowSize=5))
self.assertEqual(
sum(historicalValuesAll),
sum(estimatorParams3["movingAverage"]["historicalValues"]))
self.assertEqual(totalAll, estimatorParams3["movingAverage"]["total"])
def _generateAnomalyParams(self, metricID, statsSampleCache,
defaultAnomalyParams):
"""
Generate the model's anomaly likelihood parameters from the given sample
cache.
:param metricID: the metric ID
:param statsSampleCache: a sequence of MetricData instances that
comprise the cache of samples for the current inference result batch with
valid raw_anomaly_score in the processed order (by rowid/timestamp). At
least self._statisticsMinSampleSize samples are needed.
:param defaultAnomalyParams: the default anomaly params value; if can't
generate new ones (not enough samples in cache), this value will be
returned verbatim
:returns: new anomaly likelihood parameters; defaultAnomalyParams, if there
are not enough samples in statsSampleCache.
"""
if len(statsSampleCache) < self._statisticsMinSampleSize:
# Not enough samples in cache
# TODO: unit-test this
self._log.error(
"Not enough samples in cache to update anomaly params for model=%s: "
"have=%d, which is less than min=%d; firstRowID=%s; lastRowID=%s.",
metricID, len(statsSampleCache), self._statisticsMinSampleSize,
statsSampleCache[0].rowid if statsSampleCache else None,
statsSampleCache[-1].rowid if statsSampleCache else None)
return defaultAnomalyParams
# We have enough samples to generate anomaly params
lastRowID = statsSampleCache[-1].rowid
numSamples = min(len(statsSampleCache), self._statisticsSampleSize)
# Create input sequence for algorithms
samplesIter = itertools.islice(statsSampleCache,
len(statsSampleCache) - numSamples,
len(statsSampleCache))
scores = tuple((
row.timestamp,
row.metric_value,
row.raw_anomaly_score,
) for row in samplesIter)
assert len(scores) >= self._statisticsMinSampleSize, (
"_generateAnomalyParams: samples count=%d is smaller than min=%d; "
"model=%s; lastRowID=%s") % (
len(scores),
self._statisticsMinSampleSize,
metricID,
lastRowID,
)
assert len(scores) <= self._statisticsSampleSize, (
"_generateAnomalyParams: samples count=%d is larger than max=%d; "
"model=%s; lastRowID=%s") % (
len(scores),
self._statisticsSampleSize,
metricID,
lastRowID,
)
# Calculate estimator parameters
# We ignore statistics from the first day of data (288 records) since the
# CLA is still learning. For simplicity, this logic continues to ignore the
# first day of data even once the window starts sliding.
_, _, params = algorithms.estimateAnomalyLikelihoods(
anomalyScores=scores, skipRecords=NUM_SKIP_RECORDS)
anomalyParams = {}
anomalyParams["last_rowid_for_stats"] = lastRowID
anomalyParams["params"] = params
self._log.debug(
"Generated anomaly params for model=%s using "
"numRows=%d with rows=[%s..%s]", metricID, numSamples,
statsSampleCache[-numSamples].rowid, statsSampleCache[-1].rowid)
return anomalyParams
def _generateAnomalyParams(self, metricID, statsSampleCache,
defaultAnomalyParams):
"""
Generate the model's anomaly likelihood parameters from the given sample
cache.
:param metricID: the metric ID
:param statsSampleCache: a sequence of MetricData instances that
comprise the cache of samples for the current inference result batch with
valid raw_anomaly_score in the processed order (by rowid/timestamp). At
least self._statisticsMinSampleSize samples are needed.
:param defaultAnomalyParams: the default anomaly params value; if can't
generate new ones (not enough samples in cache), this value will be
returned verbatim
:returns: new anomaly likelihood parameters; defaultAnomalyParams, if there
are not enough samples in statsSampleCache.
"""
if len(statsSampleCache) < self._statisticsMinSampleSize:
# Not enough samples in cache
# TODO: unit-test this
self._log.error(
"Not enough samples in cache to update anomaly params for model=%s: "
"have=%d, which is less than min=%d; firstRowID=%s; lastRowID=%s.",
metricID, len(statsSampleCache), self._statisticsMinSampleSize,
statsSampleCache[0].rowid if statsSampleCache else None,
statsSampleCache[-1].rowid if statsSampleCache else None)
return defaultAnomalyParams
# We have enough samples to generate anomaly params
lastRowID = statsSampleCache[-1].rowid
numSamples = min(len(statsSampleCache), self._statisticsSampleSize)
# Create input sequence for algorithms
samplesIter = itertools.islice(
statsSampleCache,
len(statsSampleCache) - numSamples,
len(statsSampleCache))
scores = tuple(
(row.timestamp, row.metric_value, row.raw_anomaly_score,)
for row in samplesIter)
assert len(scores) >= self._statisticsMinSampleSize, (
"_generateAnomalyParams: samples count=%d is smaller than min=%d; "
"model=%s; lastRowID=%s") % (len(scores), self._statisticsMinSampleSize,
metricID, lastRowID,)
assert len(scores) <= self._statisticsSampleSize, (
"_generateAnomalyParams: samples count=%d is larger than max=%d; "
"model=%s; lastRowID=%s") % (len(scores), self._statisticsSampleSize,
metricID, lastRowID,)
# Calculate estimator parameters
# We ignore statistics from the first day of data (288 records) since the
# CLA is still learning. For simplicity, this logic continues to ignore the
# first day of data even once the window starts sliding.
_, _, params = algorithms.estimateAnomalyLikelihoods(
anomalyScores=scores, skipRecords=NUM_SKIP_RECORDS)
anomalyParams = {}
anomalyParams["last_rowid_for_stats"] = lastRowID
anomalyParams["params"] = params
self._log.debug("Generated anomaly params for model=%s using "
"numRows=%d with rows=[%s..%s]",
metricID, numSamples,
statsSampleCache[-numSamples].rowid,
statsSampleCache[-1].rowid)
return anomalyParams
