What are the tradeoffs associated with the various XGBoost interfaces?

I’m spinning up with XGBoost today, and have already encountered at least two ways to train gradient boosted trees:

Approach 1 (“native” xgb)

source: XGB python intro

xg_train = xgb.DMatrix(data = X_train, label = y_train)
xg_train = xgb.DMatrix('./data/processed/train.buffer')

param = {'max_depth':2, 'eta':1, 'objective':'binary:logistic'}
clf = xgb.train(param,xg_train, 100) # 1.79s wall time

Approach 2 (“sklearn-style”)

source: XGB sklearn wrapper example

clf2 = xgb.XGBClassifier(
    n_estimators=100, max_depth=2, eta=1, objective="binary:logistic",
clf2.fit(train[['distance_from_net', 'angle']], train['is_goal'].astype(int))

I’ve checked that the results (i.e. predictions on a validation set) are identical, but are there hidden tradeoffs associated with each approach? i.e:

  • speed? Is data binarization handled behind the scenes in approach 2? I recorded the following with timeit:
Approach 1  1.11 s ± 25.2 ms per loop (7 runs)
Approach 2  1.58 s ± 367 ms per loop (7 runs)
  • completeness do the interfaces in Approach 2 or 3 give up access to any features of “native” xgb?

The two interfaces should have identical set of features. As for speed, the sklearn wrapper automatically wraps the data in DMatrix data structure and incurs some overhead. The difference between two interfaces will shrink if you add the DMatrix construction time to the time for xgb.train().

1 Like