3

I used a Dimensionality Reduction method (discussion here: Random projection algorithm pseudo code) on a large dataset.

After reducing the dimension from 1000 to 50, I get my new dataset where each sample looks like:

[ 1751. -360. -2069. ..., 2694. -3295. -1764.]

Now I am a bit confused, because I don't know what negative feature values supposed to mean. Is it okay to have negative features like this? Because before the reduction, each sample was like this:

3, 18, 18, 18, 126 ...

Is it normal or am I doing something wrong?

Community
  • 1
  • 1
Ahmed
  • 219
  • 1
  • 3
  • 8

1 Answers1

3

I guess you implemented the algorithm from this paper.

As the projection matrix has some negative entries it is ok that the projection maps positve to negative values. So the change in the signs does not indicate an error.

rocksportrocker
  • 7,251
  • 2
  • 31
  • 48
  • 2
    Since @Ahmed prompted me to comment on this question too -- Your old vectors are some linear combination of the new basis vectors for the reduced-dimension space. Your vectors will be expressible sometimes only by using a bit of one new dimension, but subtracting out some of another. It's a bit hard to extend the "feature" interpretation. But say one item in your old data set was the band Queen, who has elements of rock and opera. Say one dimension in your new basis is "rock" and another maybe means "not like opera". Queen would be some "rock" minus some "not like opera". – Sean Owen Sep 20 '11 at 10:56