Search code examples
nlpldatopic-modelingdimensionality-reduction

Dimensionality Reduction before Topic Modeling with LDA

I want to do some topic modeling with LDA, but unfortunately my data is pretty sparse and the results are not satisfying. Because I still would like to try to solve my task with LDA (even though there might be better possibilities), I am thinking about using some kind of dimensionality reduction before LDA.

I am aware of the fact that LDA is used for topic modelling but also can be used for dimensionality reduction, so does it even make sense to try to reduce the dimensionality before using LDA? And if yes, what methods can I use? I think it wouldn’t make sense to use something like LSI or SVD.

Solution

  • As you pointed out LDA can be considered a Dimensionality Reduction technic. Hence, I would say that it does not really make sense.

    However, often LDA is used in combination with tf/idf and stop word filtering. This allows to remove too sparse and meaningless words.