Influence of tuning to CART

[Gu-Youngfeng]

In Guo et al. (ase '13), the parameters minbucket and minsplit of cart is set as follows,

S = len(train_set)
if S <= 100:
    minbucket = np.floor((S/10)+(1/2))
    minsplit = 2*minbucket
else:
    minsplit = np.floor((S/10)+(1/2))
    minbucket = np.floor(minsplit/2)

if minbucket < 2:
    minbucket = 2
if minsplit < 4:
    minsplit = 4

minbucket = int(minbucket) # cart cannot set a float minbucket or minsplit 
minsplit = int(minsplit)

cart_model = DecisionTreeRegressor( min_samples_split = minsplit, min_samples_leaf = minbucket)

But in Nair et al. (ase '17) , they simply remove the parameters of cart as follows,

cart_model = DecisionTreeRegressor()

So is there a big difference between tuned cart and non-tuned cart?

[Gu-Youngfeng]

In the experiment of performance prediction in terms of MMRE, the non-tuned cart seems performs well than tuned cart. The latter will result in large MMRE.

Compare the 2 kinds of cart among all datasets in /data/ might be a good solution.