LogLoss
Negative log-likelihood of true labels under the predicted probability distribution.
Log Loss (cross-entropy loss) penalises confident wrong predictions much more heavily than uncertain ones. Unlike accuracy or F1, it evaluates the full probability output of the classifier rather than just the argmax label, rewarding well-calibrated models.
Lower values indicate better performance (MAXIMIZE = False). A perfect
classifier achieves log loss of 0; a random classifier on a binary problem
achieves approximately ln(2) ≈ 0.693.
::
Log Loss = -(1/N) · Σᵢ Σ_c yᵢ_c · log(pᵢ_c)
where yᵢ_c is 1 if sample i belongs to class c and pᵢ_c is the predicted probability.
Range: [0, +∞), lower is better (MAXIMIZE = False).
References
Methods
score(true_labels: 'DashAIDataset', probs_pred_labels: 'np.ndarray', multiclass: Optional[bool] = None) -> float
LogLossCalculate Log Loss score between true labels and predicted labels.
Parameters
- true_labels : DashAIDataset
- A DashAI dataset with labels.
- probs_pred_labels : np.ndarray
- A two-dimensional matrix in which each column represents a class and the row values represent the probability that an example belongs to the class associated with the column.
- multiclass : bool, optional
- Whether the task is a multiclass classification. If None, it will be determined automatically from the number of unique labels.
Returns
- float
- Log Loss score between true labels and predicted labels
get_metadata(cls: 'BaseMetric') -> Dict[str, Any]
BaseMetricGet metadata values for the current metric.
Returns
- Dict[str, Any]
- Dictionary with the metadata
is_multiclass(true_labels: 'np.ndarray') -> bool
ClassificationMetricDetermine if the classification problem is multiclass (more than 2 classes).
Parameters
- true_labels : np.ndarray
- Array of true labels.
Returns
- bool
- True if the problem has more than 2 unique classes, False otherwise.
Compatible with
TabularClassificationTaskImageClassificationTaskTextClassificationTask