scikit-learn 1.6 breaks boa's MSE metric

[jemissik]

scikit-learn removed 'squared' option for mean_squared error in version 1.6, so this metric is broken in boa

[jemissik]

I tried just installing scikit-learn<1.6, but now am running into issues with ax/numpy. see facebook/Ax#2524

[jemissik]

I ran into this again while testing BOA 0.11.0 in a fresh environment.

The plain MeanSquaredError metric still works. The issue is the RMSE-style metrics that call:

mean_squared_error(..., squared=False)

I reproduced this with:

• boa-framework 0.11.0
• scikit-learn 1.8.0
• numpy 1.26.4

In scikit-learn 1.8.0, mean_squared_error has signature:

mean_squared_error(y_true, y_pred, *, sample_weight=None, multioutput="uniform_average")

so BOA's RootMeanSquaredError fails with:

TypeError: got an unexpected keyword argument 'squared'

The affected BOA code appears to be:

• boa/metrics/metrics.py: RootMeanSquaredError sets _metric_to_eval = "mean_squared_error" and injects {"squared": False}
• boa/metrics/metric_funcs.py: normalized_root_mean_squared_error() calls mean_squared_error(..., squared=False, ...)

scikit-learn deprecated the squared argument in 1.4 and removed it in 1.6. The recommended replacement is sklearn.metrics.root_mean_squared_error, which was added in scikit-learn 1.4.

Suggested fix:

from sklearn.metrics import root_mean_squared_error

Then:

• change BOA's RootMeanSquaredError to use root_mean_squared_error directly
• change normalized_root_mean_squared_error() to compute RMSE using root_mean_squared_error
• update BOA's dependency to something like scikit-learn>=1.4,<2

As a temporary workaround for users, pinning scikit-learn<1.6 avoids the breakage.