sklearn中log_loss报错：不支持多输出目标数据的标签二值化

Hey there! Let's break down why this error is popping up and how to fix it.

The Core Issue #

Sklearn's log_loss function has a default behavior that's tripping you up here: when your y_true is a 2D array, it automatically assumes you're working on a multi-output task (like predicting multiple independent labels for each sample) instead of a single multi-class classification task where each row represents the true probability distribution over classes.

Even though mathematically log_loss should handle true probability distributions, the default multi_class='auto' parameter checks the shape of y_true and misinterprets your input as multi-output data— which doesn't support the label binarization step the function tries to run in that case.

How to Fix It #

You just need to explicitly tell the function that you're dealing with a multinomial (multi-class) classification task where each row in y_true is a probability distribution over classes. Here's how:

Option 1: Specify multi_class='multinomial' (Recommended)

This approach preserves the true probability distribution information in your y_true data, which aligns perfectly with what you're trying to do:

from sklearn import metrics
import numpy as np
y_true = np.array([[0.2,0.8,0],[0.9,0.05,0.05]])
y_predict = np.array([[0.5,0.5,0.0],[0.5,0.4,0.1]])
# Explicitly define the task type to use multinomial log loss
metrics.log_loss(y_true, y_predict, multi_class='multinomial')

Option 2: Convert y_true to class indices (Not Ideal for Your Use Case)

If you didn't need to retain the true probability distributions, you could convert y_true to 1D indices of the most probable class. But this loses the nuance of your original true distribution, so it's not recommended here:

y_true_indices = np.argmax(y_true, axis=1)
metrics.log_loss(y_true_indices, y_predict)

Why This Works #

When you set multi_class='multinomial', the log_loss function switches to using the multinomial log loss formula, which is specifically designed to handle true class probability distributions as input— exactly what you're providing.

内容的提问来源于stack exchange，提问作者user1700890