EXTEND_MODEL_CASES flag in PredictTimeSeries - is it actual training on the new "extended" data ?


  • Hi to all, I have Enterprise SQL OLAP Server.

    I want to extend AND TRAIN my existing mining model with new data, all manuals including MSDN, assert it is possible with EXTEND_MODEL_CASES in PREDICTION JOIN queries.

    I build the query as described, and i receive some results indeed.

    But, when i test the mining model, to check the new cases actually added to existing cases, with this simple test query:

    select top 10 * FROM MINING STRUCTURE MineStruct1.CASES order by time desc

    i see that, there is no single new case added, all the existing cases are old, that populated with "insert into .." initial process.

    So, it looks like, the added cases is not any training, just temporary substitution, and being such, does not affects prediction algorithm noway.

    But now i see no value in such technique, for what thing it may be good ?

    Is it the only way to get use of updated data - is to reprocess the existing mining model totally, even when added only several new records ?

    I very appreciate any explanations, help, or just notes

    (Tags: DMX, processing, model, cases, on the fly, train, time series, prediction)

    Saturday, April 14, 2012 10:16 PM

All replies

  • Basically, adding the parameter in the prediction query does not retrain your model, or add to your training cases in the mining structure. You must reprocess your model to do that. And, if you have a lot of new data, you probably should retrain your model.

    However, if you don't have a lot of new values, EXTEND_MODEL_CASES takes advantage of the formula generated by the model and adjusts the predictions using the new data. You can "back up" the new values into your existing series, or tack them on the end. Just be careful to get the right start and end points for your predictions.

    Take a look at the following tutorial to see how EXTEND_MODEL_CASES works.

    The parameter REPLACE_MODEL_CASES works very differently. Here is an example in an end-to-end scenario:

    Let me know if you have any questions. I apologize for the delay in replying.


    SQL Server UE, Data Mining

    Tuesday, December 24, 2013 7:41 PM