I am trying to run the sequential K means algorithm as described here on a corpus using tf-idf as a vectorized representation of my documents. I do this because I don't have access to all of my documents at once, since I am constantly receiving new data. This method should allow me to obtain clusters gradually evolving.
The thing is, the only way to compute tf-idf for a given document is to know the whole dataset in order to be able to compute the inverse frequency of a word. When a new document shows up, it is easy to compute its corresponding vector knowing all the documents we have received so far. But this new document is as well modifying the entire corpus, and should modify all the previous vectors, hence making the algorithm impossible to run.
Is there a way do some kind of approximation to keep the current values of all the previously computed vectors ?
