[xiaoh1024]

Updating unrelated parameters triggers extra feature sampling, causing equivalent built-in and custom objectives to select different features. Skip resampling in ColSampler::SetConfig() when feature_fraction and feature_fraction_seed are unchanged.

Fixes #6053.

I used an AI agent to help with the analysis, code, tests, and PR description. I reviewed the code and description.

[jameslamb]

Thanks. The fix makes sense but I'd like to see the tests made much stricter, and I've asked a clarifying question about one of them.

Please see my recommendations in #7459 (comment) for some ideas about what I mean by "much stricter".

[jameslamb]

Suggested change

	"max_depth": 1,
	"learning_rate": 0.001,

I don't see a reason to override the default for learning_rate here, let's remove this.

And I think if you follow my other recommendations, you'll need to train deeper trees than max_depth: 1.

[jameslamb]

Sorry but I am not understanding this... how is the learning rate involved in any of this? I don't understand what the reset_parameter tests are covering here.

[jameslamb]

If the goal is just to set some parameter to a non-default value to be sure the initial SetConfig() call(s) do not exit early, that makes sense to me... but the tests aren't ever checking that learning_rate = 0.001 made it through into the Booster.

Either way, please do add some assertions here checking that other overrides (learning_rate is fine) do make it through into the Booster on the C++ side. The test code could call Booster._get_loaded_param() for help with that:

LightGBM/python-package/lightgbm/basic.py

Lines 3854 to 3866 in 9474d04

	def _get_loaded_param(self) -> Dict[str, Any]:
	buffer_len = 1 << 20
	tmp_out_len = ctypes.c_int64(0)
	string_buffer = ctypes.create_string_buffer(buffer_len)
	ptr_string_buffer = ctypes.c_char_p(ctypes.addressof(string_buffer))
	_safe_call(
	_LIB.LGBM_BoosterGetLoadedParam(
	self._handle,
	ctypes.c_int64(buffer_len),
	ctypes.byref(tmp_out_len),
	ptr_string_buffer,
	)
	)

[xiaoh1024]

@jameslamb Thanks for your suggestions! reset_parameter was intended to check that resetting an unrelated parameter does not advance the feature-sampling state. I originally reset the learning rate to its existing value to keep training equivalent to the reference, making differences caused by extra sampling detectable. I have also adopted your other suggestions, just like #7459.