My tabular data has following attributes:
FEATURES : Feature A (with 5 classes) & Feature B (with 8 classes) + some continuous + some more categorical features.
TARGET var : Price
I used embeddings with default sizes for all categorical variables and trained a model, now I wanted to use these embeddings to find similarity between those objects using embeddings, for that I was wondering if I should :
A. Concatenate Embeddings, i.e. : item embeddings = [feature A emb] + [feature B emb] and then look at cosine similarity of these concatenated vectors (I am inclined to try this first)
2. create feature C ( with 5x8 = 40 classes) and retrain model with potentially bigger embedding space and then use it directly to find similar objects.
Have a gut feeling that there is a better method that i am missing… would really appreciate any help on this.