Matt,

You don't need predicted values for the anomaly likelihood.  You just need
the timestamps, actual value, and anomaly scores. A scalar "actual value"
doesn't exist for GPS but this is only used to detect the degenerate case
of completely flat values - for the gps situation you can just pass in
random values I think.

--Subutai

On Thu, Jan 29, 2015 at 1:55 PM, Matthew Taylor <[email protected]> wrote:

> This is probably a question for Chetan, but I thought I'd send it to
> the list in case anyone else ran into the same thing later.
>
> Our current geospatial demo app [1] displays the raw anomaly score,
> not the anomaly likelihood. I was looking into using the anomaly
> likelihood to see if it improves results, but the anomaly likelihood
> code requires a predicted value to work. It seems that the models
> using a CoordinateEncoder do not produce predictions in the model
> result objects, even if the "inferenceType" specified in the model
> params is "TemporalMultiStep".
>
> So it seems like I can't calculate the likelihood value or get the
> predicted locations from the model. Is this a known deficiency or
> something I'm doing wrong?
>
> [1] https://github.com/numenta/nupic.geospatial
>
> Thanks,
> ---------
> Matt Taylor
> OS Community Flag-Bearer
> Numenta
>
>

Reply via email to