Perda Logarítmica, also known as logistic loss or perda de entropia cruzada, is a performance metric used primarily in classificação binária problems where the predicted output is a probability value ranging between 0 and 1. It quantifies the difference between the predicted probabilities and the actual class labels (0 or 1). The objective of using log loss is to evaluate how well a classification model predicts probabilities for binary outcomes, with lower values indicating better desempenho do modelo.
Matematicamente, a Log Loss é calculada usando a seguinte fórmula:
Log Loss = -1/N * Σ [y * log(p) + (1 – y) * log(1 – p)]
Onde:
- N é o número total de previsões.
- y é o rótulo real (0 ou 1).
- p é a probabilidade prevista da classe positiva (1).
The Log Loss value ranges from 0 to infinity, where 0 indicates perfect predictions (model outputs probabilities of either 0 or 1 correctly) and larger values indicate worse performance. A model that predicts probabilities close to the true labels will have a lower log loss, while a model that predicts probabilities far das etiquetas verdadeiras incorrerá em uma perda de log mais alta.
Log Loss is particularly useful in scenarios where the output is not just a hard classification but a probability, making it suitable for applications such as regressão logística, neural networks, and other probabilistic classifiers. It is also widely usada em aprendizado de máquina competições, como as realizadas pelo Kaggle, para avaliar o desempenho do modelo.