Exporting Triple Exponential Smoothing Model

[#1]

Does anyone know if there is a way to export a Triple Exponential Smoothing model?

In Online Mode, I was not able to get HANA Triple Exponential Smoothing to allow me to save the model, which normally means I can export it to PMML or a stored procedure in HANA. I was not able to try, yet, HANA R-Triple Exponential Smoothing due to not having Rserve installed (any quick guides would be appreciated).

In Offline Mode, I am able to successfully run Triple Exponential Smoothing, but I get an error when I click "Save as Model". The error, which follows, seems to indicate that export this model is not supported (...then why is it an option...?):

R-Triple Exponential Smoothing works in Offline Mode as well, but there is no "Save as Model" option at all available.

So, I have one option left to try, but I figured it would be best to just ask the community if I am missing something or if this is just not possible in Predictive Analysis with this algorithm?

[ashokkumar_kn · Accepted solution]

Hi Kevin,

For Time Series we do not have model.

Always it works on new records and predicts future in time.

May be there is UX improvement needed, we will hide the option save as model insted of message.

Thanks,

Ashok

[Former Member · Accepted solution]

Hi Ashok,

Thanks for letting me know this. Are there plans to support exporting time series models?

Is there a complete list of what algorithms can be exported, in which formats, and whether or not they only work in online or offline mode?

Yeah the UI is a bit confusing with how some options are or are not available and they they may or may not work. when you get errors it's sometimes unclear what needs fixed. For example, for online mode, I had to find out on SCN that Triple Exponential Smoothing only works with a two column view, even though it lets you select said model after connecting to a table that doesn't match this requirement.

thanks!

[Former Member · Accepted solution]

Hi Kevin,

In regards to you question on a complete list of what algorithms can be exported.

Here is a link to a list of machine learning algorithms - all of which you should be able to train on known data and export to use on unknown data:

http://en.wikipedia.org/wiki/List_of_machine_learning_algorithms

(there is a few exceptions - for instance the k-means algorithm isn't a model that you would train, save and run on unknown data).

As for the "forecasting" algorithms that you mentioned in the first message such as 1-2-3 exponential smoothing if you look at the equation - it is purely based on previous time-periods & values and as such not a learning model.

Hope this helps you on your way.

Best regards,

Kurt Holst

[Former Member · Accepted solution]

Kurt--

I really don't understand this statement: "for instance the k-means algorithm isn't a model that you would train, save and run on unknown data".  One of the most common applications of clustering algorithms is customer segmentation which is EXACTLY a situation where you would want to segment your existing customers using predictors and then validate that the segmentation algorithm against purchase/profit metrics and then apply the model to new customers or prospects to evaluate the predicted purchasing habits at the beginning of the relationship.

K-means is an algorithm that is easily scored on other data, but there is no built in R function to do it (an odd oversight, I think).  It is certainly something that should exist and isn't nearly as difficult to implement as many other algorithms.  I frequently score new data by writing a manual scoring calculation in SQL that uses the cluster centers table output from R.  SAS provides scoring functionality in its k-means algorithm PROC FASTCLUS.

[Former Member · Accepted solution]

Hi Hillary,

Ups, you are absolutely right. My mistake!

Thanks for bringing this to my attention.

Of course the clustering algorithms like k-means are trainable - how could I forget the famous IRIS example - hope my statistics teacher isn't reading this blog.

So looking at the list on wikipedia it looks like all of the algorithms are trainable and hence could be used in a train -> save -> run on unknown data scenario.

http://en.wikipedia.org/wiki/List_of_machine_learning_algorithms

Best regards,

Kurt Holst