`scale_pos_weight` and `sample_weight` give inconsistent results.

[j-at-ch]

Setting the model parameter scale_pos_weight and the fit parameter sample_weight are giving inconsistent results (not due to randomness) while fitting binary classifiers using the sklearn interface.

Language: Python 3.10.12 XGBoost Version: 2.1.1

Printout from MWE:

> class_weight: {0: 1.0, 1: 4.1} scale_pos_weight: 4.1 sample_weights: [1.  4.1]
> accuracy using scale_pos_weight (mean, std, n):  (0.625, 0.0)
> accuracy using class_weight (mean, std, n):  (0.635, 0.0)

MWE :

import numpy as np
import xgboost as xgb
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from sklearn.utils import class_weight

np.random.seed(42)

# Synthetic data
X = np.random.rand(1000, 10)
y = np.zeros(1000)

# Introduce class imbalance (e.g., 80% class 0, 20% class 1)
y[800:] = 1

# Split data into train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Calculate scale_pos_weight for XGBoost
spw = np.round(np.sum(y_train == 0) / np.sum(y_train == 1), 1)
cws = {0: 1.0, 1: spw}

sample_weight = class_weight.compute_sample_weight(
    class_weight=cws,
    y=y_train
)

print(f'class_weight: {cws}, scale_pos_weight: {spw}, sample_weights: {np.unique(sample_weight)}')

def train_and_evaluate(n_runs, model_params, fit_params):
    accuracy_list = []
    for run in range(n_runs):
        mod = xgb.XGBClassifier(**model_params)
        mod.fit(X_train, y_train, **fit_params)
        y_pred = mod.predict(X_test)
        accuracy = accuracy_score(y_test, y_pred)
        accuracy_list.append(accuracy)
    return np.mean(accuracy_list), np.std(accuracy_list)

# Set the number of repeat runs
n_runs = 5

# Train XGBoost with scale_pos_weight
model_params = {'scale_pos_weight': spw, 'random_state': 42, 'n_estimators': 7}
fit_params = {}
mean_acc, std_acc = train_and_evaluate(n_runs, model_params, fit_params)
print(f'accuracy using scale_pos_weight (mean, std, n):  {mean_acc, std_acc}')

# Train XGBoost with class_weight
model_params = {'random_state': 42, 'n_estimators': 7}
fit_params = {'sample_weight': sample_weight}
mean_acc, std_acc = train_and_evaluate(n_runs, model_params, fit_params)
print(f'accuracy using class_weight (mean, std, n):  {mean_acc, std_acc}')