Common Loss Functions for Classification

While regression tasks often utilize loss functions like Mean Squared Error or Mean Absolute Error, classification problems require different measures to quantify the discrepancy between predicted class probabilities (or scores) and the actual discrete class labels. For Gradient Boosting Machines (GBMs), the choice of loss function is significant because its negative gradient defines the "pseudo-residuals" that subsequent base learners are trained to predict.

Let's examine the most common loss functions used for classification within the GBM framework.

Binary Classification: Log Loss (Binomial Deviance)

For binary classification problems, where the target variable $y$ takes values in $\{0, 1\}$, the standard and most widely used loss function is the Log Loss, also known as Binomial Log-Likelihood or Binomial Deviance.

It stems directly from the principle of maximum likelihood estimation. If $F(x)$ is the raw output (logit or log-odds) predicted by the model for an input $x$, the probability of the positive class ($y=1$) is typically obtained via the logistic (sigmoid) function:

$$p(x) = P(y=1 | x) = \frac{1}{1 + e^{-F(x)}}$$

The Log Loss for a single observation $(x, y)$ is defined as the negative logarithm of the likelihood of the true label given the predicted probability:

$$L(y, p(x)) = -[y \log(p(x)) + (1-y) \log(1 - p(x))]$$

This loss function heavily penalizes predictions that are confidently wrong. For instance, if the true label is $y=1$ but the predicted probability $p(x)$ is close to 0, the term $-y \log(p(x)) = -\log(p(x))$ approaches infinity. Conversely, if the prediction is correct and confident ($p(x)$ close to 1 when $y=1$), the loss approaches 0.

Binary Log Loss when the true label is 1. The loss increases significantly as the predicted probability approaches 0 (incorrect prediction).

Gradient and Pseudo-Residuals: In the GBM algorithm, we need the negative gradient of the loss function with respect to the model's raw output $F(x)$ at the current iteration $m-1$, denoted $F_{m-1}(x)$. This gradient guides the fitting of the next tree $h_m(x)$.

$$-\frac{\partial L(y, F(x))}{\partial F(x)} = -\frac{\partial L}{\partial p} \frac{\partial p}{\partial F(x)}$$

Using the chain rule, and knowing that $\frac{\partial p}{\partial F(x)} = p(x)(1-p(x))$ (the derivative of the sigmoid function) and $\frac{\partial L}{\partial p} = -\frac{y}{p} + \frac{1-y}{1-p} = \frac{p-y}{p(1-p)}$, we get:

$$-\frac{\partial L(y, F(x))}{\partial F(x)} = - \left( \frac{p(x) - y}{p(x)(1-p(x))} \right) (p(x)(1-p(x))) = y - p(x)$$

This result is remarkably simple. The negative gradient, which serves as the pseudo-residual $r_{im}$ for observation $i$ at iteration $m$, is simply the difference between the true label (0 or 1) and the currently predicted probability $p_{i, m-1} = P(y_i=1 | x_i; F_{m-1})$.

$$r_{im} = y_i - p_{i, m-1}$$

Therefore, each new tree $h_m(x)$ in a GBM trained with Log Loss for binary classification is fitted to predict the residual error of the current probability estimate.

Multi-Class Classification: Log Loss (Multinomial Deviance)

The concept extends naturally to multi-class classification problems where the target variable $y$ can belong to one of $K$ classes, $\{1, ..., K\}$. We typically use one-hot encoding for the true labels, so $y_i$ is a vector where $y_{ik}=1$ if observation $i$ belongs to class $k$, and $y_{ik}=0$ otherwise.

The model needs to output $K$ scores or logits, $F_1(x), ..., F_K(x)$. These are converted into probabilities $p_k(x) = P(y=k | x)$ using the softmax function:

$$p_k(x) = \frac{e^{F_k(x)}}{\sum_{j=1}^K e^{F_j(x)}}$$

The loss function used is the Multinomial Log Loss, also known as Multinomial Deviance or Cross-Entropy Loss:

$$L(\{y_k\}, \{p_k(x)\}) = - \sum_{k=1}^K y_k \log(p_k(x))$$

Since only one $y_k$ is 1 and the others are 0 for a given observation, this simplifies to $-\log(p_c(x))$, where $c$ is the true class index.

Gradient and Pseudo-Residuals: Similar to the binary case, we compute the negative gradient of the loss with respect to each raw output $F_k(x)$. The calculation is slightly more involved due to the softmax function, but the result remains elegant. For a specific class $k$ and observation $i$, the negative gradient at iteration $m-1$ is:

$$r_{imk} = -\frac{\partial L(y_i, F_{m-1}(x_i))}{\partial F_{k, m-1}(x_i)} = y_{ik} - p_{ik, m-1}$$

Where $p_{ik, m-1}$ is the predicted probability for class $k$ for observation $i$ based on the model $F_{m-1}$.

Again, the pseudo-residual for each class is the difference between the true label (0 or 1 in the one-hot encoding) and the currently predicted probability for that class.

In practice, most GBM implementations handle multi-class problems by building $K$ separate regression trees at each boosting iteration $m$. The $k$-th tree is trained using the pseudo-residuals $r_{imk} = y_{ik} - p_{ik, m-1}$ as its target variable. The outputs of these $K$ trees are then combined to update the overall model scores $F_k(x)$ for each class.

Why Log Loss?

Log Loss (or Multinomial Deviance) is preferred over alternatives like Exponential Loss (used in the original AdaBoost algorithm) primarily because it is less sensitive to outliers or mislabeled data points. Exponential Loss penalizes errors exponentially, meaning a single very wrong prediction can dominate the gradient calculation and potentially distort the model. Log Loss provides a more robust measure of performance for classification tasks, directly related to information theory (cross-entropy) and probabilistic interpretation (log-likelihood). Its smooth gradients also facilitate the optimization process within the GBM framework.

By understanding how these classification loss functions operate and how their gradients yield the pseudo-residuals, we gain insight into the core mechanism driving the learning process in Gradient Boosting Machines for classification problems. The choice directly influences what errors the subsequent trees attempt to correct, shaping the final predictive model.