Implementation:Online ml River Metrics CrossEntropy
| Knowledge Sources | |
|---|---|
| Domains | Online_Learning, Evaluation_Metrics |
| Last Updated | 2026-02-08 16:00 GMT |
Overview
Multi-class cross-entropy loss metric, a generalization of logarithmic loss for multiple classes.
Description
CrossEntropy computes the cross-entropy loss between predicted probability distributions and true labels. It measures how well predicted probabilities match the true distribution, with lower values indicating better predictions. The metric requires probability distributions as predictions (dictionaries mapping class labels to probabilities) rather than hard class labels.
Usage
Use CrossEntropy to evaluate probabilistic classifiers that output class probability distributions. Unlike metrics that require hard labels, CrossEntropy evaluates the quality of probability estimates, penalizing confident wrong predictions more heavily. It's particularly useful for multi-class classification where you want to assess prediction confidence calibration.
Code Reference
Source Location
- Repository: Online_ml_River
- File: river/metrics/cross_entropy.py
Signature
class CrossEntropy(metrics.base.MeanMetric, metrics.base.MultiClassMetric):
def __init__(self, cm=None):
pass
Import
from river import metrics
I/O Contract
| Method | Parameters | Returns | Description |
|---|---|---|---|
| update | y_true (label), y_pred (dict) | None | Updates with true label and probability distribution |
| get | - | float | Returns average cross-entropy loss (lower is better) |
Note: requires_labels property returns False (requires probability distributions).
Usage Examples
from river import metrics
y_true = [0, 1, 2, 2]
y_pred = [
{0: 0.29450637, 1: 0.34216758, 2: 0.36332605},
{0: 0.21290077, 1: 0.32728332, 2: 0.45981591},
{0: 0.42860913, 1: 0.33380113, 2: 0.23758974},
{0: 0.44941979, 1: 0.32962558, 2: 0.22095463}
]
metric = metrics.CrossEntropy()
for yt, yp in zip(y_true, y_pred):
metric.update(yt, yp)
print(metric.get())
# 1.222454
# 1.169691
# 1.258864
# 1.321597
print(metric)
# CrossEntropy: 1.321598
# Lower cross-entropy indicates better probability estimates
# Perfect predictions would have cross-entropy approaching 0