Fix SR anomaly score calculation at beginning (#5502)

* adjust expected value

* update boundary calculation

* fix boundary

* adjust default values

* fix percent case

* fix error in anomaly score calculation

* adjust score calculation for first & second points

* fix sr do not report anomaly at beginning

* fix a issue in batch process

* remove a unused parameter

Co-authored-by: yuyi@microsoft.com <Yuanxiang.Ying@microsoft.com>

src/Microsoft.ML.TimeSeries/SrCnnEntireAnomalyDetector.cs

@@ -309,6 +309,15 @@ public void Process()
                     _previousBatch = _previousBatch.GetRange(_batch.Count, _bLen);
                     _previousBatch.AddRange(_batch);
                     _modeler.Train(_previousBatch.ToArray(), ref _results);
+
+                    // move the values to front
+                    for (int i = 0; i < _batch.Count; ++i)
+                    {
+                        for (int j = 0; j < _outputLength; ++j)
+                        {
+                            _results[i][j] = _results[_bLen + i][j];
+                        }
+                    }
                 }
                 else
                 {
@@ -334,7 +343,7 @@ public ValueGetter<VBuffer<double>> CreateGetter(DataViewRowCursor input, string
                         double src = default;
                         srcGetter(ref src);
                         var result = VBufferEditor.Create(ref dst, _outputLength);
-                        _results[input.Position % _batchSize + _bLen].CopyTo(result.Values);
+                        _results[input.Position % _batchSize].CopyTo(result.Values);
                         dst = result.Commit();
                     };
                 return getter;
@@ -351,6 +360,15 @@ internal sealed class SrCnnEntireModeler
             private static readonly double _deanomalyThreshold = 0.35;
             private static readonly double _boundSensitivity = 93.0;
             private static readonly double _unitForZero = 0.3;
+            private static readonly double _minimumScore = 0.0;
+            private static readonly double _maximumScore = 1.0;
+            // If the score window is smaller than this value, the anomaly score is tend to be small.
+            // Proof: For each point, the SR anomaly score is calculated as (w is average window size):
+            // (mag - avg_mag) / avg_mag
+            // = max (w * mag_{a} - sum_{i=0 to w-1} mag_{a - i}) / sum_{i=0 to w-1} mag_{a - i}
+            // = max ((w - 1) * mag_{a} + C) / (mag_{a} + C)
+            // <= w - 1
+            private static readonly int _minimumScoreWindowSize = (int)(_maximumScore * 10) + 1;
 
             //    pseudo-code to generate the factors.
             //    factors = []
@@ -577,15 +595,20 @@ private void SpectralResidual(double[] values, double[][] results, double thresh
                 {
                     _ifftMagList[i] = Math.Sqrt(_ifftRe[i] * _ifftRe[i] + _ifftIm[i] * _ifftIm[i]);
                 }
+
                 AverageFilter(_ifftMagList, Math.Min(_ifftMagList.Length, _judgementWindowSize));
+                for (int i = 0; i <= Math.Min(length, _minimumScoreWindowSize); ++i)
+                {
+                    _cumSumList[i] = _cumSumList[Math.Min(length, _minimumScoreWindowSize) - 1];
+                }
 
                 // Step 7: Calculate raw score and set result
                 for (int i = 0; i < results.GetLength(0); ++i)
                 {
                     var score = CalculateScore(_ifftMagList[i], _cumSumList[i]);
                     score /= 10.0f;
-                    score = Math.Min(score, 1);
-                    score = Math.Max(score, 0);
+                    score = Math.Min(score, _maximumScore);
+                    score = Math.Max(score, _minimumScore);
 
                     var detres = score > threshold ? 1 : 0;
 

test/Microsoft.ML.TimeSeries.Tests/TimeSeriesDirectApi.cs

@@ -776,6 +776,64 @@ public void TestSrCnnAnomalyDetectorWithSeasonalAnomalyData(
             }
         }
 
+        [Theory, CombinatorialData]
+        public void TestSrCnnAnomalyDetectorWithAnomalyAtBeginning(
+            [CombinatorialValues(SrCnnDeseasonalityMode.Stl, SrCnnDeseasonalityMode.Mean, SrCnnDeseasonalityMode.Median)] SrCnnDeseasonalityMode mode
+        )
+        {
+            var ml = new MLContext(1);
+            IDataView dataView;
+            List<TimeSeriesDataDouble> data;
+
+            var dataPath = GetDataPath("Timeseries", "anomaly_at_beginning.csv");
+
+            // Load data from file into the dataView
+            dataView = ml.Data.LoadFromTextFile<TimeSeriesDataDouble>(dataPath, hasHeader: true);
+            data = ml.Data.CreateEnumerable<TimeSeriesDataDouble>(dataView, reuseRowObject: false).ToList();
+
+            // Setup the detection arguments
+            string outputColumnName = nameof(SrCnnAnomalyDetection.Prediction);
+            string inputColumnName = nameof(TimeSeriesDataDouble.Value);
+
+            // Do batch anomaly detection
+            var options = new SrCnnEntireAnomalyDetectorOptions()
+            {
+                Threshold = 0.30,
+                BatchSize = -1,
+                Sensitivity = 80.0,
+                DetectMode = SrCnnDetectMode.AnomalyAndMargin,
+                Period = 0,
+                DeseasonalityMode = mode
+            };
+
+            var outputDataView = ml.AnomalyDetection.DetectEntireAnomalyBySrCnn(dataView, outputColumnName, inputColumnName, options);
+
+            // Getting the data of the newly created column as an IEnumerable of SrCnnAnomalyDetection.
+            var predictionColumn = ml.Data.CreateEnumerable<SrCnnAnomalyDetection>(
+                outputDataView, reuseRowObject: false);
+
+            var anomalyIndex = 1;
+
+            int k = 0;
+            foreach (var prediction in predictionColumn)
+            {
+                Assert.Equal(7, prediction.Prediction.Length);
+                if (anomalyIndex == k)
+                {
+                    Assert.Equal(1, prediction.Prediction[0]);
+                    Assert.True(prediction.Prediction[6] > data[k].Value || data[k].Value > prediction.Prediction[5]);
+                }
+                else
+                {
+                    Assert.Equal(0, prediction.Prediction[0]);
+                    Assert.True(prediction.Prediction[6] <= data[k].Value);
+                    Assert.True(data[k].Value <= prediction.Prediction[5]);
+                }
+
+                ++k;
+            }
+        }
+
         [Theory, CombinatorialData]
         public void TestSrcnnEntireDetectNonnegativeData(
             [CombinatorialValues(true, false)] bool isPositive)

test/data/Timeseries/anomaly_at_beginning.csv

@@ -0,0 +1,39 @@
+Value
+181.944
+37.176
+57.14
+67.128
+72.12
+77.112
+82.104
+83.1
+87.09
+92.088
+92.01
+97.08
+102.072
+107.05
+107.06
+117.048
+122.04
+132.024
+147
+151.82
+151.992
+151.72
+151.94
+156.969
+156.984
+156.92
+161.976
+161.94
+161.97
+166.968
+176.952
+181.94
+186.936
+201.91
+201.912
+201.9
+206.904
+216.88