Mapping of anomaly-types on time domain

[timeeval]

Mapping of anomaly-types on time domain

[CodeLionX]

Dear @janthmueller,

Unfortunately, there is no automatic way to get the anomaly type annotations for already generated time series when there are multiple types defined. GutenTAG chooses a random position (with some constraints) for each anomaly in the specified region ("beginning", "middle", "end"); they are not sorted. You should be able to visually tell them apart, though.

However, I think you could modify GutenTAG to include the anomaly types in the datasets and re-generate the benchmark datasets. You won't be able to get the same datasets, though, because we generated the benchmark datasets very early on, and GutenTAG's reproducibility features (TS-name-based RNG seeding and writing out the seed and GutenTAG version) were introduced later on.
We have some students working with GutenTAG that need to do something similar. I will contact them and ask how they do it.

[CodeLionX]

Our students do the following to encode the anomaly type into the "is_anomaly" column:

diff --git a/gutenTAG/anomalies/__init__.py b/gutenTAG/anomalies/__init__.py
index 7c82b51..7aa6fc3 100644
--- a/gutenTAG/anomalies/__init__.py
+++ b/gutenTAG/anomalies/__init__.py
@@ -58,11 +58,12 @@ class Anomaly:
             start, end = self.exact_position, self.exact_position + self.anomaly_length
 
         length = end - start
-        label_range = LabelRange(start, length)
+        label_range = LabelRange(start, length, "")
         protocol = AnomalyProtocol(start, end, self.channel, ctx, label_range, creep_length=self.creep_length)
 
         for anomaly in self.anomaly_kinds:
             protocol = anomaly.generate(protocol)
+            protocol.labels.class_label = type(anomaly).__name__
 
         return protocol
 
diff --git a/gutenTAG/anomalies/types/__init__.py b/gutenTAG/anomalies/types/__init__.py
index ddc3b85..8236963 100644
--- a/gutenTAG/anomalies/types/__init__.py
+++ b/gutenTAG/anomalies/types/__init__.py
@@ -13,6 +13,7 @@ from ...utils.types import AnomalyGenerationContext
 class LabelRange:
     start: int
     length: int
+    class_label: str
 
 
 @dataclass
diff --git a/gutenTAG/base_oscillations/utils/consolidator.py b/gutenTAG/base_oscillations/utils/consolidator.py
index 3708632..3287e44 100644
--- a/gutenTAG/base_oscillations/utils/consolidator.py
+++ b/gutenTAG/base_oscillations/utils/consolidator.py
@@ -68,6 +68,28 @@ class Consolidator:
     def _add_label_ranges_to_labels(self, label_ranges: List[LabelRange]):
         if self.labels is not None:
             for label_range in label_ranges:
-                self.labels[label_range.start:label_range.start + label_range.length] = 1
+                if label_range.class_label == "AnomalyAmplitude":
+                    nr_label_anomaly = 1
+                elif label_range.class_label == "AnomalyExtremum":
+                    nr_label_anomaly = 2
+                elif label_range.class_label == "AnomalyFrequency":
+                    nr_label_anomaly = 3
+                elif label_range.class_label == "AnomalyMean":
+                    nr_label_anomaly = 4
+                elif label_range.class_label == "AnomalyPattern":
+                    nr_label_anomaly = 5
+                elif label_range.class_label == "AnomalyPatternShift":
+                    nr_label_anomaly = 6
+                elif label_range.class_label == "AnomalyPlatform":
+                    nr_label_anomaly = 7
+                elif label_range.class_label == "AnomalyTrend":
+                    nr_label_anomaly = 8
+                elif label_range.class_label == "AnomalyVariance":
+                    nr_label_anomaly = 9
+                elif label_range.class_label == "AnomalyModeCorrelation":
+                    nr_label_anomaly = 10
+                else:
+                    nr_label_anomaly = -1
+                self.labels[label_range.start:label_range.start + label_range.length] = nr_label_anomaly
         else:
             raise AssertionError("You cannot run this method before initializing the `labels` field!")

You can then use the anomaly category number to figure out the used anomaly kind.

Warning

Please keep in mind that this just uses the last used anomaly kind as the final anomaly type. GutenTAG allows users to specify multiple anomaly kinds per anomaly and stacks them. So you have to make sure that your configurations don't use this feature!

[janthmueller]

Thank you for the detailed answer. I will probably implement it within the next few days similar to your students. Nevertheless, it would be nice to see it as a feature within the application in the future. Including the labeling of time steps of overlapping and differing anomaly types.

[CodeLionX]

We don't see this as one of the main features of GutenTAG, but we are open for community contributions. If you are already implementing it for yourself, you could also start working on a PR.

Please discuss architectural changes with us before submitting a PR so that we can agree on it.