|
I was looking at LDA from a perspective of it fitting other datasets than words. When would the assumption that allow a dataset to be used with the LDA break down, like for example if I have a dataset with each document replaced by a person and the words replaced by scores/abilities in various subjects? Assuming I have a trained LDA model if we present it with a datapoint m that has only a few of the values of the features(terms), if I do the inference and get the phi(p(terms|hidden=k)) and the theta(p(hidden|datapoint = m)). Would I be correct in doing the following operations First per new datapoint calculate: p(terms) = sum_topic(p(terms|topic)*p(topic|document=m)) Now I believe the p(terms) should now represent the probability of terms most likely with probability values filled in for the missing features(terms). What is wrong with such and assumption? |
