使用不同参数返回相同分数的LightGBM模型

Question

我正在尝试在Kaggle Iowa住房数据集上训练一个LightGBM模型，我写了一个小脚本，在给定的范围内随机尝试不同的参数。我不确定我的代码出了什么问题，但是脚本用不同的参数返回相同的分数，这是不应该发生的。我在Catboost上尝试了相同的脚本，它可以正常工作，所以我猜问题出在LGBM上。

代码：

import numpy as np
import pandas as pd
from lightgbm import LGBMRegressor
from sklearn.metrics import mean_absolute_error
from sklearn.model_selection import train_test_split
from random import choice, randrange, uniform 

complete_train = pd.read_csv(
    "train.csv",
    encoding = "UTF-8", 
    index_col = "Id")

complete_test = pd.read_csv(
    "test.csv",
    encoding = "UTF-8",
    index_col = "Id")

def encode_impute(*datasets):
    for dataset in datasets:
        for column in dataset.columns:
            dataset[
                column].fillna(
                -999,
                inplace = True)
            if dataset[
                column].dtype ==  "object":
                dataset[
                    column] = dataset[
                    column].astype("category", copy = False)
encode_impute(
    complete_train, 
    complete_test)

X = complete_train.drop(
    columns = "SalePrice")

y = complete_train[
    "SalePrice"]

X_train, X_valid, y_train, y_valid = train_test_split(X, y)

def objective():

    while True:

        params = {
            "boosting_type": choice(["gbdt", "goss", "dart", "rf"]),
            "num_leaves": randrange(10000),
            "learning_rate": uniform(0.01, 1),
            "subsample_for_bin": randrange(100000000),
            "min_data_in_leaf": randrange(100000000),
            "reg_alpha": uniform(0, 1),
            "reg_lambda": uniform(0, 1),
            "feature_fraction": uniform(0, 1),
            "bagging_fraction": uniform(0, 1),
            "bagging_freq": randrange(1, 100)}

        params["bagging_fraction"] = 1.0 if params[
                        "boosting_type"] == "goss" else params[
                        "bagging_fraction"]

        model = LGBMRegressor().set_params(**params)

        model.fit(X_train, y_train)

        predictions = model.predict(X_valid)

        error_rate = mean_absolute_error(
          y_valid, predictions)

        print(f"Score = {error_rate} with parameters: {params}","\n" *5)

objective()

我得到的输出示例：

分数= 55967.70375930444，参数：{'boosting_type'：'gbdt'，'num_leaves'：6455，'learning_rate'：0.2479700848039991，'subsample_for_bin'：83737077，'min_data_in_leaf'：51951103，'reg_alpha'：0.1856001984332697，'reg_lambda'：0.7849262049058852，'feature_fraction'：0.10550627738309537，'bagging_fraction'：0.2613298736131875，'bagging_freq'：96}

分数= 55967.70375930444，参数：{'boosting_type'：'dart'，'num_leaves'：9678，'learning_rate'：0.28670432435369037，'subsample_for_bin'：24246091，'min_data_in_leaf'：559094，'reg_alpha'：0.07261459695501371，'reg_lambda'：0.8834743560240725，'feature_fraction'：0.5361519020265366，'bagging_freq'：0.9120030047714073，‘bagging_freq’：10}

分数= 55967.70375930444，参数：{'boosting_type'：'goss'，'num_leaves'：4898，'learning_rate'：0.09237499846487345，'subsample_for_bin'：32620066，'min_data_in_leaf'：71317820，'reg_alpha'：0.9818297737748625，'reg_lambda'：0.11638265354331834，'feature_fraction'：0.4230342728468828，'bagging_fraction'：1.0，‘bin_freq’：64}

Answer 1

我要指出的是，所有选项中的min_data_in_leaf参数似乎都很高，我怀疑模型没有学习到任何东西，只是发送只有根节点的响应变量的平均值。