Support NaN values in UpliftTree and UpliftRandomForest

[aman-coder03]

This PR adds native support for missing values (NaNs) in UpliftTree and
UpliftRandomForest.

Each candidate split evaluates both possible NaN routing directions
(left/right) and learns the optimal routing per node, similar to
scikit-learn’s decision tree behavior.

The learned NaN routing is stored in each tree node and applied
consistently during training, pruning, filling, and prediction.

This resolves #802.

[CLAassistant]

All committers have signed the CLA.

[aman-coder03]

The Read the Docs build failure appears to be due to Cython extensions requiring a compiler in the RTD environment.
No documentation-related files were modified in this PR.
Happy to adjust the RTD config if you’d like me to handle this

[jeongyoonlee]

Thanks for your contribution, @aman-coder03. I left a few comments. Can you address them? Also, please add test code to tests/test_uplift_trees.py accordingly. Thanks!

[jeongyoonlee]

We need this code block. Is there any reason you removed it?

[jeongyoonlee]

We need this code block for early stopping.

[jeongyoonlee]

NaN handling is needed here in the percentile calculation.

[aman-coder03]

@jeongyoonlee i will look into this!

[aman-coder03]

hey @jeongyoonlee i have implemented the changes, please have a look!

[jeongyoonlee]

@aman-coder03, I see some code blocks necessary (e.g., code for the CTS evaluation criteria) are removed without any comments. Also, some tests are failing.

[aman-coder03]

Thanks for pointing that out. I will restore the original CTS and evaluation blocks and limit changes strictly to NaN routing logic to preserve original behavior @jeongyoonlee

[aman-coder03]

@jeongyoonlee can you please have a look at the code changes now

[jeongyoonlee]

Code review

Found 1 issue:

1. np.isnan() called unconditionally on non-numeric columns in divideSet, causing a TypeError for any dataset with string/categorical features.

The new NaN routing code in divideSet applies np.isnan(X[:, column]) without checking whether the column is numeric:

causalml/causalml/inference/tree/uplift.pyx

Lines 916 to 921 in 4e84b3f

	filt = X[:, column] == value
	# Handle NaNs
	nan_mask = np.isnan(X[:, column])
	if missing_go_to_left:

divideSet handles both numeric splits (isinstance(value, numbers.Number)) and string splits (the else branch). When X[:, column] is an object array of strings, np.isnan raises:

TypeError: ufunc 'isnan' not supported for the input types

The companion function divideSet_len in this same PR correctly guards the NaN logic with if np.issubdtype(X[:, column].dtype, np.number)::

causalml/causalml/inference/tree/uplift.pyx

Lines 960 to 967 in 4e84b3f

	# Handle NaNs only for numeric columns
	if np.issubdtype(X[:, column].dtype, np.number):
	nan_mask = np.isnan(X[:, column])
	if missing_go_to_left:
	filt = filt | nan_mask
	else:

The same unguarded np.isnan pattern also appears in group_counts_by_divide:

causalml/causalml/inference/tree/uplift.pyx

Lines 316 to 321 in 4e84b3f

	# Handle NaNs
	cdef np.ndarray[np.uint8_t, ndim=1, cast=True] nan_mask = np.isnan(col_vals)
	if missing_go_to_left:
	filt = filt | nan_mask

The fix is to wrap both np.isnan calls with the same numeric dtype guard used in divideSet_len. The test added in this PR only injects NaNs into a numeric feature column, so it would not catch this regression.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}