Skip to content

Latest commit

 

History

History
79 lines (64 loc) · 2.24 KB

README.md

File metadata and controls

79 lines (64 loc) · 2.24 KB
Raw

Torsk
Citekey HeimAvery2019Adaptive
Source Code https://github.com/nmheim/torsk
Learning type unsupervised
Input dimensionality multivariate

Dependencies

  • python 3
  • joblib
  • numpy
  • pandas
  • matplotlib
  • scipy
  • scikit-learn

Notes

  • Returns window anomaly scores. We can transform them back to point scores using post-processing:
from timeeval.utils.window import ReverseWindowing
# post-processing for Torsk
def _post_torsk(scores: np.ndarray, args: dict) -> np.ndarray:
    pred_size = args.get("hyper_params", {}).get("prediction_window_size", 20)
    context_window_size = args.get("hyper_params", {}).get("context_window_size", 10)
    size = pred_size * context_window_size + 1
    return ReverseWindowing(window_size=size).fit_transform(scores)

  • We don't use the input mapping feature to its full potential. Only a single random weight mapping is used. The parameters of this mapping are exposed.

  • The pytorch implementation had various errors, so we disabled this backend and removed the code for it.

  • Information about the different window sizes and why the actual window size is one larger than train_size + pred_size:

    dataset_size = 40
train_size = 20
transient_size = 10
pred_size  = 10
steps = dataset_size - train_size - 1 - pred_size = 9

------------------------------------------
|          dataset size = 40             |
------------------------------------------

For training:
++++++++++++++++++++++++++++++++++
|         window size = 31       |
|       train 21      |  pred 10 |
|     train_x 20     |    -      |
||      train_y 20    |    -     |
++++++++++++++++++++++++++++++++++

For optimization:
++++++++++++++++++++++++++++++++++
|         window size = 31       |
|  tran 10 | opt data |     -    |
++++++++++++++++++++++++++++++++++

For prediction:
++++++++++++++++++++++++++++++++++
|         window size = 31       |
|          -          |  pred 10 |
++++++++++++++++++++++++++++++++++

  • Parameter values are very dependent on a dataset. The default parameters are just rough estimates.