Training crashes when "monotone_constraints_method" is set to "advanced"

[sv1990]

Description

Training crashes when "monotone_constraints_method" is set to "advanced" and at least one feature is categorical but "monotone_constraints" is not defined.

The crash is fixed by setting "monotone_constraints", but according to the documentation "monotone_constraints_method" should not be used unless "monotone_constraints" is set.

Reproducible example

import lightgbm as lgb
import numpy as np
import pandas as pd

size = 60
df = pd.DataFrame(
    {
        "x": np.random.uniform(size=size),
        "c": pd.Categorical(np.random.choice([0, 1, 2], size=size, replace=True)),
        "y": np.random.uniform(size=size),
    }
)

lgb.train(
    {
        "monotone_constraints_method": "advanced",
        # "monotone_constraints": "0,0",
    },
    lgb.Dataset(df[["x", "c"]], df["y"], categorical_feature="auto"),
)

Environment info

LightGBM version or commit hash: '4.6.0'

Command(s) you used to install LightGBM

poetry add lightgbm

Additional Comments

[jameslamb]

Thanks for using LightGBM, we'll look into this soon.

To clarify... does "crashes" mean something like the process exits, a segmentation fault, etc.? Or do you just mean that it raises a Python exception?

Are there any logs printed before it "crashes", and if so could you share those too?

[sv1990]

This is the output I get from the example above:

[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.000430 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 25
[LightGBM] [Info] Number of data points in the train set: 60, number of used features: 2
[LightGBM] [Info] Start training from score 0.505615
fish: Job 1, 'python test.py' terminated by signal SIGSEGV (Address boundary error)