This is an exercise in both improving my Haskell and better understanding how decision tree learning works –- but mostly the former.
Apart from split ordering, there is no attempt to make things efficient and there is currently no pruning or other standard tricks to improve the learning performance.
The src/run.hs
file is what sets up, runs, and evaluates a simple decision
tree learner on a data set supplied via the command line.
To build this program, use:
cabal build run
This will create an executable in dist/build/run/run
. I usually make a
symbolic link to it from the top level directory:
ln -s dist/build/run/run run
Then a tree can be trained and evaluated like so:
./run data/winequality-red.csv
The resulting utput shows the tree that was built and its mean square error on the traing set.
The R library rpart
can also be used to build and evaluate decision trees.
For comparison, here is how to build a tree using rpart
:
library(rpart)
wine <- read.csv('data/winequality-red.csv',sep=';')
fit <- rpart(quality ~ ., data = wine)
summary(fit)