search

regel / loudml

Loud ML is the first open-source AI solution for ICT and IoT automation
Other
293 stars 92 forks source link

Filling missing feature values by last known value #26

Closed regel closed 5 years ago

regel commented 6 years ago

1.3.0

When defining features in a time series model, missing values will be NaN and prevent the model from giving predictions.

This occurs, eg when using mean(parameter) aggregation and data points are missing in a given time bucket, causing the mean to be undefined at that point in time.

With 1.3.0, the 'default' property gives the ability to fill these NaN (missing values), with a static default value, eg 0.

Default values could cause false positives in anomaly detection, therefore another possibility is to replicate the last know value for a given feature, if the value is missing.

regel commented 5 years ago

The feature data-model is augmented in 1.4 with 'default': 'previous', implemented in 1.4.0 release.