使用稀疏数据，训练 LDA（潜在狄利克雷分配）并预测新文档的更快方法是什么？

Question

1. About training LDA :

   When we try to implement LDA, we need to construct a words-documents matrix, but this matrix is sparse data, because our token dictionary is very large(some millions words ), and a documents has a small set of token (~ 1000 - 10000 words), so in words-documents matrix, it has so many 0 values, and it will take longer time to train model. So how can we do faster ?

2. About predict a new documents :

   After training, now we have a new LDA model, so we can use it to predict which topics with a new documents. But before feed new documents to our model, we need to convert it to a words vector, and its vector length will be our dictionary length ( some millions words). So, it will have some many zeros values, and in fact, our cost time is increased by vector length.

So is documents-words matrix a effective way to implement LDA ? Can we have other better way? I need some recommend for my project , so please help

Answer 1

With sparse data, you should of course use sparse vectors instead of dense vectors.

Instead of storing all the zeros, you only keep the non-zero values.

A typical data model (see literature for alternatives) is to simply use a list of tuples (i,v) where i is the column index, and v is a non-zero value.