weighted topic modeling. Experiments in augmenting topic-modeling methods, if given an a priori measurement of the distances between words

Given the newsgroup dataset, construct a low-dimensional topic embedding. For each newsgroup, construct a mean vector from all documents in it. Evaluation is the percent of documents that are closest (in embedding space) to their correct label.

This seems to help both LDA and LSA methods, except for low-dimension LDA embeddings. Notably, the best results on this benchmark came from LSA with tf-idf weights, with the improved method.

The method forces parts of these "sparse" algorithms to be dense. This can be augmented by only storing the closest-N pairwise distances for each word. However, this would still make the procedure more dense.

There is also a small subset of the newsgroup dataset in here. Along with a script to fetch GLoVe vectors to generating embedding distances. There is also a version of Funk's SVD that I implemented to do approximate LSA, but this didn't really work well