Custom binary_crossentropy that penalizes one mispredicted class

[vabatista]

Hi, I have a dataset with 95% of data within one class and 5% of another. My models are having 94% of accuracy, but the model predicts 100% the prevalent class. My best roc_auc_score is about 0.71 on validation set.

Can anyone suggest a good custom objetive function that solves this problem?

[ChristianThomae]

You may use class_weight with inverse class frequencies when calling fit, see the docs.

[Yingyingzhang15]

@ChristianThomae Can you give an example of how to use the class-weight? I mean do I need to pass an array or something else to the fitfunction?

[tboquet]

You could also take a look at #2115 if you want an example of a loss function where you can weight the different cases of a confusion matrix.

[Yingyingzhang15]

@tboquet Thank you! It seems that I have solved this problem.The class weight should be a dictionary as explained in docs.