Options
Building a model
Hello everyone,
I have some questions on building predictive models.
First of all, if I know for sure that some of the future values I want to predict will be zero, how could I handle it?
Secondly, if I fulfill my dataset with many factors that affects my attributes, is there any way to give the % for every factor that affects the attribute?
Finally, I sometimes have as results negative values, but in reality I can't have negative values, only positive.
Is there any way to solve this problem?
Thanks in advance.
I have some questions on building predictive models.
First of all, if I know for sure that some of the future values I want to predict will be zero, how could I handle it?
Secondly, if I fulfill my dataset with many factors that affects my attributes, is there any way to give the % for every factor that affects the attribute?
Finally, I sometimes have as results negative values, but in reality I can't have negative values, only positive.
Is there any way to solve this problem?
Thanks in advance.
0
Best Answer

OptionsTelcontar120 Moderator, RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 1,635 UnicornHi @Papad,
 If you know for certain that you want to predict zero under certain conditions, and you don't trust your learner to understand those conditions because they aren't represented sufficiently in your training set, then the best thing for you to do is to write a separate rule (IF...THEN logic using Generate Attributes) to handle those cases.
 If you are looking for univariate relationships between your attributes and your label, you can use the Weight by operators such as Weight by Correlation for numerical relationships and Weight by Information Gain or Weight by Chi Squared for categorical relationships. If you are talking about the contribution within a multivariate model, things are more complicated. Some of the operators output variable importance either in the form of coefficients (think standardized coefficients in regression models) or in variable lists. However other models have very nonlinear, localized predictions and for those models you might want to look at the Explain Predictions operator, which gives localized approximations for variable importance given a certain set of inputs.
 I think this question is the same as the other thread that we are already discussingplease see my latest comments there.
1