Problem: Strange loss and metric progression for focal loss.

When training with focal loss the loss actually increases and the MCC
metric gets worse over iterations when looking at it in tensorboard.

For log loss it works but for focal loss it looks off and I would suspect that alpha of 0.5
and a gamma close to zero should give similar results to log loss.

If i use the the learning rate found by the logloss using model the problem persists.

If I reduce the learning rate to a very low level and increase the gamma it works

With a low gamma and a low learning rate the problem is back.

Focal Loss seems to need a different range for the learning rate than logloss

catboost version: 1.2.2

Here is some code to reproduce this finding:

from itertools import product

import numpy as np
from catboost import CatBoostClassifier, Pool, cv
from sklearn.metrics import roc_auc_score
from sklearn.datasets import make_blobs

from matplotlib import pyplot
from pandas import DataFrame

generate 2d classification dataset

X,y = make_blobs(n_samples=1000, centers=2, n_features=1000)

Split dataset into train and test sets

train_size = int(0.8 * len(X))
X_train, X_test = X[:train_size], X[train_size:]
y_train, y_test = y[:train_size], y[train_size:]

Create CatBoost dataset

train_pool = Pool(X_train, y_train)

#train a model using log loss

loss = "Logloss"
model_log_loss = CatBoostClassifier(loss_function=loss, eval_metric="MCC", verbose=0)
test_pool = Pool(X_test)

model_log_loss.fit(train_pool)

#now train a model using focal loss
#with learning rate from ther log loss model
#alpha and gamma were set to make it close
#to log loss

loss = "Focal:focal_alpha=0.5;focal_gamma=0.0001"
model_focal = CatBoostClassifier(learning_rate = model_log_loss.learning_rate_, loss_function=loss, eval_metric="MCC", verbose=0)

model_focal.fit(train_pool)

#also trained a model with a very low learning rate and a higher gamma this seems to fix it

but focal loss seems to need a much lower leanring rate than log loss

loss = "Focal:focal_alpha=0.5;focal_gamma=3"
model_focal_low_learning = CatBoostClassifier(learning_rate = 0.00001, loss_function=loss, eval_metric="MCC", verbose=0)

model_focal_low_learning.fit(train_pool)

#also trained a model with a very low learning rate and low gamma then the problem shows up again

loss = "Focal:focal_alpha=0.5;focal_gamma=0.00001"
model_focal_low_learning = CatBoostClassifier(learning_rate = 0.00001, loss_function=loss, eval_metric="MCC", verbose=0)