From e351ecc7eecd8dab0ef761be462aba751f0bc1e0 Mon Sep 17 00:00:00 2001 From: zeotuan <48720253+zeotuan@users.noreply.github.com> Date: Mon, 11 Mar 2024 09:47:05 +1100 Subject: [PATCH] Anomaly Detection: Add Daily Season with Hourly Interval to HoltWinter (#546) * Add Daily Season with Hourly Interval, Add custom periodicity constructor * Add Custom seriesPeriodicity, Hourly interval with Daily seasonality tests * Add Test using custom seriesPeriodicity --- .../seasonal/HoltWinters.scala | 43 ++++++------ .../seasonal/HoltWintersTest.scala | 65 +++++++++++++++++++ 2 files changed, 87 insertions(+), 21 deletions(-) diff --git a/src/main/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWinters.scala b/src/main/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWinters.scala index 0ee0ac25f..203dfe9fa 100644 --- a/src/main/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWinters.scala +++ b/src/main/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWinters.scala @@ -25,11 +25,11 @@ import collection.mutable.ListBuffer object HoltWinters { object SeriesSeasonality extends Enumeration { - val Weekly, Yearly: Value = Value + val Daily, Weekly, Yearly: Value = Value } object MetricInterval extends Enumeration { - val Daily, Monthly: Value = Value + val Hourly, Daily, Monthly: Value = Value } private[seasonal] case class ModelResults( @@ -48,29 +48,30 @@ object HoltWinters { } -/** - * Detects anomalies based on additive Holt-Winters model. The methods has two - * parameters, one for the metric frequency, as in how often the metric of interest - * is computed (e.g. daily) and one for the expected metric seasonality which - * defines the longest cycle in series. This quantity is also referred to as periodicity. - * - * For example, if a metric is produced daily and repeats itself every Monday, then the - * model should be created with a Daily metric interval and a Weekly seasonality parameter. - * - * @param metricsInterval: How often a metric is available - * @param seasonality: Cycle length (or periodicity) of the metric - */ -class HoltWinters( - metricsInterval: HoltWinters.MetricInterval.Value, - seasonality: HoltWinters.SeriesSeasonality.Value) +class HoltWinters(seriesPeriodicity: Int) extends AnomalyDetectionStrategy { import HoltWinters._ - private val seriesPeriodicity = seasonality -> metricsInterval match { - case (SeriesSeasonality.Weekly, MetricInterval.Daily) => 7 - case (SeriesSeasonality.Yearly, MetricInterval.Monthly) => 12 - } + /** + * Detects anomalies based on additive Holt-Winters model. The methods has two + * parameters, one for the metric frequency, as in how often the metric of interest + * is computed (e.g. daily) and one for the expected metric seasonality which + * defines the longest cycle in series. This quantity is also referred to as periodicity. + * + * For example, if a metric is produced daily and repeats itself every Monday, then the + * model should be created with a Daily metric interval and a Weekly seasonality parameter. + * + * @param metricsInterval : How often a metric is available + * @param seasonality : Cycle length (or periodicity) of the metric + */ + def this(metricsInterval: HoltWinters.MetricInterval.Value, + seasonality: HoltWinters.SeriesSeasonality.Value) = + this(seasonality -> metricsInterval match { + case (HoltWinters.SeriesSeasonality.Daily, HoltWinters.MetricInterval.Hourly) => 24 + case (HoltWinters.SeriesSeasonality.Weekly, HoltWinters.MetricInterval.Daily) => 7 + case (HoltWinters.SeriesSeasonality.Yearly, HoltWinters.MetricInterval.Monthly) => 12 + }) /** * Triple exponential smoothing with additive trend and seasonality diff --git a/src/test/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWintersTest.scala b/src/test/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWintersTest.scala index decf5a91c..36689180b 100644 --- a/src/test/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWintersTest.scala +++ b/src/test/scala/com/amazon/deequ/anomalydetection/seasonal/HoltWintersTest.scala @@ -207,6 +207,71 @@ class HoltWintersTest extends AnyWordSpec with Matchers { anomalies should have size 3 } } + + "work on hourly data with daily seasonality" in { + // https://www.kaggle.com/datasets/fedesoriano/traffic-prediction-dataset + val hourlyTrafficData = Vector[Double]( + 15, 13, 10, 7, 9, 6, 9, 8, 11, 12, 15, 17, 16, 15, 16, 12, 12, 16, 17, 20, 17, 19, 20, 15, + 14, 12, 14, 12, 12, 11, 13, 14, 12, 22, 32, 31, 35, 26, 34, 30, 27, 27, 24, 26, 29, 32, 30, 27, + 21, 18, 19, 13, 11, 11, 11, 14, 15, 29, 33, 32, 32, 29, 27, 26, 28, 26, 25, 29, 26, 24, 25, 20, + 18, 18, 13, 13, 10, 12, 13, 11, 13, 22, 26, 27, 31, 24, 23, 26, 26, 24, 23, 25, 26, 24, 26, 24, + 19, 20, 18, 13, 13, 9, 12, 12, 15, 16, 23, 24, 25, 24, 26, 22, 20, 20, 22, 26, 22, 21, 21, 21, + 16, 18, 19, 14, 12, 13, 14, 14, 13, 20, 22, 26, 26, 21, 23, 23, 19, 19, 20, 24, 18, 19, 16, 17, + 16, 16, 10, 9, 8, 7, 9, 8, 12, 13, 17, 14, 14, 14, 14, 11, 15, 13, 12, 17, 18, 17, 16, 15, 13 + ) + + val strategy = new HoltWinters( + HoltWinters.MetricInterval.Hourly, + HoltWinters.SeriesSeasonality.Daily) + + val nDaysTrain = 6 + val nDaysTest = 1 + val trainSize = nDaysTrain * 24 + val testSize = nDaysTest * 24 + val nTotal = trainSize + testSize + + val anomalies = strategy.detect( + hourlyTrafficData.take(nTotal), + trainSize -> nTotal + ) + + anomalies should have size 2 + } + + "work on monthly data with yearly seasonality using custom seriesPeriodicity" in { + // https://datamarket.com/data/set/22ox/monthly-milk-production-pounds-per-cow-jan-62-dec-75 + val monthlyMilkProduction = Vector[Double]( + 589, 561, 640, 656, 727, 697, 640, 599, 568, 577, 553, 582, + 600, 566, 653, 673, 742, 716, 660, 617, 583, 587, 565, 598, + 628, 618, 688, 705, 770, 736, 678, 639, 604, 611, 594, 634, + 658, 622, 709, 722, 782, 756, 702, 653, 615, 621, 602, 635, + 677, 635, 736, 755, 811, 798, 735, 697, 661, 667, 645, 688, + 713, 667, 762, 784, 837, 817, 767, 722, 681, 687, 660, 698, + 717, 696, 775, 796, 858, 826, 783, 740, 701, 706, 677, 711, + 734, 690, 785, 805, 871, 845, 801, 764, 725, 723, 690, 734, + 750, 707, 807, 824, 886, 859, 819, 783, 740, 747, 711, 751, + 804, 756, 860, 878, 942, 913, 869, 834, 790, 800, 763, 800, + 826, 799, 890, 900, 961, 935, 894, 855, 809, 810, 766, 805, + 821, 773, 883, 898, 957, 924, 881, 837, 784, 791, 760, 802, + 828, 778, 889, 902, 969, 947, 908, 867, 815, 812, 773, 813, + 834, 782, 892, 903, 966, 937, 896, 858, 817, 827, 797, 843 + ) + + val strategy = new HoltWinters(12) + + val nYearsTrain = 3 + val nYearsTest = 1 + val trainSize = nYearsTrain * 12 + val testSize = nYearsTest * 12 + val nTotal = trainSize + testSize + + val anomalies = strategy.detect( + monthlyMilkProduction.take(nTotal), + trainSize -> nTotal + ) + + anomalies should have size 7 + } } object HoltWintersTest {