feat: fallback to stddev if threshold is too low

Signed-off-by: Avik Basu <[email protected]>

numalogic/models/threshold/_median.py

@@ -5,6 +5,9 @@
 from numalogic.base import BaseThresholdModel
 from numalogic.tools.exceptions import InvalidDataShapeError, ModelInitializationError
 
+import logging
+
+LOGGER = logging.getLogger(__name__)
 _INLIER: Final[int] = 0
 _OUTLIER: Final[int] = 1
 _INPUT_DIMS: Final[int] = 2
@@ -19,18 +22,20 @@ class MaxPercentileThreshold(BaseThresholdModel):
  min_threshold: Value to be used if threshold is less than this
  """
 
- __slots__ = ("_max_percentile", "_min_thresh", "_thresh", "_is_fitted")
+ __slots__ = ("_max_percentile", "_min_thresh", "_thresh", "_is_fitted", "_adjust_threshold")
 
  def __init__(
  self,
  max_inlier_percentile: float = 96.0,
  min_threshold: float = 1e-4,
+ adjust_threshold: bool = False,
  ):
  super().__init__()
  self._max_percentile = max_inlier_percentile
  self._min_thresh = min_threshold
  self._thresh = None
  self._is_fitted = False
+ self._adjust_threshold = adjust_threshold
 
  @property
  def threshold(self):
@@ -45,6 +50,17 @@ def _validate_input(x: npt.NDArray[float]) -> None:
  def fit(self, x: npt.NDArray[float]) -> Self:
  self._validate_input(x)
  self._thresh = np.percentile(x, self._max_percentile, axis=0)
+
+ if self._adjust_threshold:
+ for idx, _ in enumerate(self._thresh):
+ if self._thresh[idx] / self._min_thresh < 1e-2:
+ LOGGER.info(
+ "Min threshold is less than 1e-2 times the "
+ "threshold for column %s; Using mean instead.",
+ idx,
+ )
+ self._thresh[idx] = np.mean(x[:, idx]) + (3 * np.std(x[:, idx]))
+
  self._thresh[self._thresh < self._min_thresh] = self._min_thresh
  self._is_fitted = True
  return self

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "numalogic"
-version = "0.9.1a3"
+version = "0.9.1a4"
 description = "Collection of operational Machine Learning models and tools."
 authors = ["Numalogic Developers"]
 packages = [{ include = "numalogic" }]