Custom loss in XGBoost is not updating

Question

Context

I am trying to use a custom loss function for an XGBoost binary classifier.

The idea was to implement in XGBoost the soft-Fbeta loss, which I read about here. Simply put: instead of using the standard logloss, use a loss function that directly optimises the Fbeta score.

Caveat

Of course, the Fbeta itself is not differentiable, so it can't be used straight out of the box. However, the idea is to use the probabilities (hence, before thresholding) to create some sort of continuous TP, FP and FN. Find more details in the referenced Medium article.

Attempt

My attempt was the following (inspired by few different people).

import numpy as np
import xgboost as xgb

def gradient(y: np.array, p: np.array, beta: float):

    """Compute the gradient of the loss function. y is the true label, p
    the probability predicted by the model """
    
    # Define the denominator
    D = p.sum() + beta**2 * y.sum() 
    
    # Compute the gradient
    grad = (1 + beta**2) * y / D - (1 + beta**2) * (np.dot(p, y)) / D**2 
        
    return grad

def hessian(y: np.array, p: np.array, beta: float):

    """Compute the Hessian of the loss function. y is the true label, p
    the probability predicted by the model """
    
    # Define the denominator
    D = p.sum() + beta**2 * y.sum() 
    
    # Tensor sum y_i + y_j
    tensor_sum = y + y[:, None]
    
    # Compute the hessian
    hess = (1 + beta**2) / D**2 * (-tensor_sum + 2*np.dot(p, y) / D)
    
    return hess

def f_smooth_loss(beta: float):
    
    """ Custom loss function for maximising F score"""
    def custom_loss(y: np.array, p: np.array):
                
        # Actual custom loss
        b = beta
        
        # Compute grad
        grad = - gradient(y, p, b)
        
        # Compute hessian
        hess = - hessian(y, p, b)
                  
        return grad, hess
        
    return custom_loss

# Random train dataset
X_train = np.random.rand(100, 100)
y_train = np.random.randint(0, 2, 100)

# Random validation dataset
X_validation = np.random.rand(1000, 100)
y_validation = np.random.randint(0, 2, 1000)

# Define a classifier trying to maximise F5 score
model = xgb.XGBClassifier(objective=f_smooth_loss(5))

# Fit
model.fit(X_train, y_train,  eval_set=[(X_train, y_train), (X_validation, y_validation)])

Output

The model runs, but the output is apparently stuck, no matter what:

[0] validation_0-logloss:0.69315    validation_1-logloss:0.69315
[1] validation_0-logloss:0.69315    validation_1-logloss:0.69315
[2] validation_0-logloss:0.69315    validation_1-logloss:0.69315
[3] validation_0-logloss:0.69315    validation_1-logloss:0.69315

Comments

1. It is possible my derivatives are not correct, even though I double checked them. However, even changing the grad and hess to constant numbers, nothing changes.

2. The Hessian here is a matrix (which would be its mathematical definition), but I think XGBoost expects a 1D array (I think it is the diagonal). However, because of point 1., nothing changes even if I change it to a 1d-array

3. Essentially, this model always predicts zeros, and does not update at all.

4. Changing the size of the (fake) dataset does not lead to any change in the logloss (even more, the numbers are exactly the same).

5. Curiously, the logloss is the same in the validation and train, this being yet another signal that there is something deeply wrong somewhere.

6. If I switch to the standard logloss (built-in), it updates (outputs are random, as the dataset is random).

Question

What is wrong in my implementation? XGB docs are pretty hard to decipher, and I can't really tell if I am missing a simple building block here.

Answer 1

The problem is that following the docs the custom loss function need the following parameters as input:

....


def f_smooth_loss(beta: float):
    
    """ Custom loss function for maximising F score"""
    def custom_loss(
        predt: np.ndarray,
        dtrain: xgb.DMatrix
    ) -> Tuple[np.ndarray, np.ndarray]:
                
        # Actual custom loss
        b = beta
        
        # Compute grad
        grad = - gradient(dtrain, predt, b)
        
        # Compute hessian
        hess = - hessian(dtrain, predt, b)
                  
        return grad, hess
        
    return custom_los

Update: following the documentation referenced about it seems that you need to pass the function in the .train() of the class not when initializing the model, e.g.:

xgb.train({'tree_method': 'hist', 'seed': 1994},  # any other tree method is fine.
           dtrain=dtrain,
           num_boost_round=10,
           obj=f_smooth_loss(5))

Also, notice that the .fit() method is a wrapper that XGBoost has as a interface to interact with other sklearn objects (e.g. sklearn.pipeline) so it might lack this functionality, so it's better to use the native method .train().

Answer 2

Please change the classifier from objective=f_smooth_loss(5) to scoring=f_smooth_loss(5):

model = xgb.XGBClassifier(scoring = f_smooth_loss(5))