negative distance

[mese79]

Hi
I just wonder is dot product enough for distance/similarity measure?
dot product for normalized features could be in range of [-1, +1], but i think two features with dot product of -1 are as dissimilar as +1 value but in different direction. So is taking the absolute distance between features a good idea?

[HobbitLong]

I am not sure if I understand correctly. But if both vectors are normalized, the absolute l2 euclidean distance is the same as cosine distance. On the other hand, if you are taking about normalized vectors, it will lead to value explosion while optimizing the contrastive loss.