anthonymorast / cppart Goto Github PK

Run test data down some trees.

lmean and rmean in giniSplit in rpart are using 1-nclasses for their mult, in cppart we used the class number

for wine this is 2 4 5 6 and 7 or something like that, likely the cause for splits going opposite of rpart occasionally.

Only do PCA on the training data (after centering the data [demeaning])
- first check if using the corerlation matrix centers the data around the origin

Translate the Training data with the principal components

How does rpart determine if it's classification or regression?

High improvement <=> low SS for anova, switch to only using the improvement for the delayed instead of SS (saves lots of computation) (rpart only uses improvement).

Same for gini, just need to find highest improvement.