topic description

for example, there are often a large number of userid and itemid, in recommendation systems that seem to directly calculate the vector, of itemid, such as the calculation in gensim, or bow or skipgram, in neural networks, but their id is a type of id.
or, for example, the method used by youtube, the vector of userid and itemid is directly spliced, and then trained directly with multi-layer dnn, but the trained userid and itemie cannot directly calculate the similarity.

so is there a way to directly train the similarity between userid and itemid and calculate the similarity?