search

rfechner / generic-hmm

Apache License 2.0
2 stars 1 forks source link

How do we cope with new states in the observation data? #7

Closed rfechner closed 2 years ago

rfechner commented 2 years ago

Since we are using a label encoder to encode data, this may become a big problem.

I think in any case, a state that was never observed before should not have an impact on the prediction. This means, that we have to encode the states of each marker of a sequence one by one and simply set the prob of observation to 0.

First, we could check if the mapping is successful. In this case, continue, else do the hacky thing of encoding one by one.

rfechner commented 2 years ago

PS this isssue should be adressed further at the end of the production, since this may hide potential errors in the code.

rfechner commented 2 years ago

Currently, we simply throw an error message containing information about which state in which marker was unseen. Alternatively, we could omit the marker from the calculation.