I am training a NN model. The data is highly imbalanced (3% for positive labels), and I have not resampled more true classes in the training set. The model performs much better when categorical cross-entropy is used.
My hypothesis is categorical cross-entropy requires 2 logits rather than 1. So the network does not have to suppress the logit for the true class that much when the model sees samples with false classes.
Does my hypothesis makes sense, and is there any conclusion on this topic? E.g., which one is better for a highly imbalanced binary classification dataset.