Can someone please help me understand how will the Cross Entropy be sensitive to the background CTR?

I didn’t understand how is NCE not sensitive to background CTR.

Hello @Vinit1

Basically, the Normalized Cross-Entropy (NCE) is equivalent to the average log-loss per impression divided by the average log-loss per impression if a model predicted the background click-through rate (CTR) for every impression.

It is not sensitive to the background CTR because the key benefit normalized cross-entropy is its robustness to the practical training of the background CTR. The least the value we get, the better is the prediction made by the model. This normalization is because the closer the background CTR to 0 or 1, the more straightforward it is to gain a better log loss.

I hope I have answered your query; please let me know if you still have any confusion.

Thank You :blush: