Why does using UpliftRandomForestClassifier consume so much memory？

[longweiwei]

Describe the bug

Train the model using UpliftRandomForestClassifier and using default configuration, When the amount of training data reaches 4.5G（about 2.9 million rows of data, 440 features）, the required memory exceeds 80G. I'm not sure if this is a bug,

To Reproduce
Can be reproduced the situation using random data。

from causalml.inference.tree import UpliftRandomForestClassifier
from causalml.metrics import plot_gain
from sklearn.model_selection import train_test_split
import causalml
causalml.__version__

from causalml.inference.tree import UpliftTreeClassifier
import numpy as np
from sys import getsizeof
import pandas as pd

treatment = np.random.randint(0, 2, (2900000)).astype(np.float32)
y =  np.random.randint(0, 2, (2900000)).astype(np.float32)
feats = np.random.randn(2900000, 440).astype(np.float32)
treatment_bak = ['treat' if v==1 else 'control' for v in treatment.tolist()]

uplift_model = UpliftRandomForestClassifier(control_name='control')
uplift_model.fit(X=feats,
                 treatment=np.array(treatment_bak),
                 y=y)

Expected behavior
Is there any way to reduce memory usage besides reducing the amount of data?

Screenshots
If applicable, add screenshots to help explain your problem.

Environment (please complete the following information):

• OS: ubuntu
• Python Version: 3.6,
• Versions of Major Dependencies (pandas, scikit-learn, cython): pandas== 1.0.4 , scikit-learn==0.24.2, cython==0 0.29.24, causalml=0.12.3

Additional context
Add any other context about the problem here.

[longweiwei]

@Pawka @mway @darthtrevino @shannonmoeller
hi! can you help with the issue

[zyq11223]

because the data type of the treatment variable is str object , it cost too much memory, and the library does not support int type

[jbbqqf]

Hi! Triaging older issues — I think this one can be closed unless someone has a fresh reproducer that isolates a real leak.

The reproducer in the body (2.9M rows × 440 features × UpliftRandomForestClassifier with default n_estimators) is at the edge of what's feasible in-memory regardless of implementation — every tree holds per-treatment-group statistics for every node, so peak RSS in the tens-of-GB range is expected at this scale, not a leak. @zyq11223's note that treatment was passed as str may also have inflated the working set; the library's int-array path (treatment_idx in causalml/inference/tree/uplift.pyx) is the cheap one.

If anyone can show a delta where memory grows monotonically across predict calls (i.e. a true leak rather than a one-shot peak), that'd be a different issue worth reopening. Otherwise the practical guidance is "subsample / fewer estimators / smaller max_depth."

(Disclosure: I drafted this comment with help from Claude Code while triaging stale issues; the references above were verified manually against the current master branch.)