[Video 3] [Discussion] I think that in TF-IDF SVD negative values should be representing abstract concepts rather than the words themselves

(Aseem Bansal) #1

In Video 3 around 28:00 we are talking about how it is hard to think about what negative values mean. I was actually thinking it is not. I understand that the matrix factorization algorithm is giving the topics like Pheneus or whatever the names were. But we cannot consider the algorithm to know that it is a name. I think it would be more about the concept. Like in Pride and Prejudice we may get Darcy as a topic. But my thoughts are that the algorithm would not be representing that as a person rather a characteristic like Darcy represents an abstract concept which may represent Pride among a few other things.

I understand that TF-IDF is just based on term frequency but isn’t matrix factorization about the main components?