Preprocess the TF-IDF representations before TruncatedSVD

[Kairixir]

Issues with sparse and dense matrices may arrise. Watch that

Possible preprocessing may be:

Pair-wise, efficient if matrix sparse:

• Chi-Squared test
• Mutual information

Other options:

• Feature agglomeration
• Random projection

If the matrix at the end of preprocessing is dense it is better to go with IncrementalPCA