Hello,
I plan performing sequence/temporal learning using one of Mahout's
classifier (probably OnlineLogisticRegression, since I need
probabilities as outcome).
For example, we have this data:
t X Y
---------------------------
T1 xT1 Y1
T2 xT2 Y2
T3 xT3 Y3
T4 xT4 Y4
where, for example xT1 is the value of X in time T1 and we know that Y
depends conditionaly on X (this means that Y is target variable and Y is
a predictor).
If we assume, that we have past-window of 2 and future-window of 1, we
can construct our training examples this way (using sliding windows)
Feature1 Feature2 TargetVariable
----------------------------------------------
xT1 xT2 Y3
xT2 xT3 Y4
...
(this learning model is based on
http://neuroph.sourceforge.net/tutorials/ChickenPricePredictionTutorial.htm)
So we get a logistic regression model for with number of features = past
window size. But what if Y depends on more than one predictor (for
example Z)? How can we map X and Z to one time unit - should we extract
also two more features for Z, analog to X)?
The second question is - what if I have size of future window >=2 ? The
problem is, that I can have only one target variable in a logistic
regression, so I can't train a model with such window size. Is that correct?
Thanks and sorry for the long post.
Svetlomir.
