Neural Networks negative prediction

svennisvenni Member Posts: 3 Contributor I
Hi guys.

I am new to this forum and fairly new to data mining.  I have been trying to use neural network to do some numerical prediction and the results should be from 0 and positive number to the thousands.  The model I have set up is predicting negative numbers as well.  How can I set it up so it only predicts positive numbers? 

Sorry if this is the wrong section to pose the question.



  • Options
    MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn

    you cannot restrict the output of the Neural Network - it is (in theory) always between positive and negative infinity.

    To tackle the problem you should first try to improve your model (optimize parameters, or try different regression algorithms). Maybe your results improve. Then you should apply some post-processing. There you have a number of options: simply dismiss negative numbers (and predict nothing), set negative predictions to zero, ... This can be done with the Generate Attributes operator.
  • Options
    svennisvenni Member Posts: 3 Contributor I
    Thanks for reply.

    I should try to explain my problem.  This is for a commodity and companies submit a price and capabilities to produce the said commodity.  A production is awarded depending on the price and the capabilities (and other factors which are not in the model).

    The data is not Linear as I was hoping so I was trying out Neural Nets for it.  My training data is previous months with the awarded production plus the price and capabilities.  And I was trying to predict the volume and what possible effect it will do to change the price.  As you can see there can never be negative volume only 0 and positive numbers. 

    Fairly new to data mining so I am probably modelling this all wrong. 

    Thanks again
  • Options
    wesselwessel Member Posts: 537 Maven
    Log transform?
  • Options
    svennisvenni Member Posts: 3 Contributor I
    Thanks for the replies.

    I did use the transformed regression for my model.  It did do the trick.  Now its up to me to play around with the model and the data.
Sign In or Register to comment.