Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.
|
|
|
|
Re: applying dimensionality reduction in clustering / classification
Posted:
Dec 12, 2012 12:48 AM
|
|
On Dec 11, 12:25 pm, tongyuent...@gmail.com wrote:
> What is the main difference when applying dimensionality reduction in classification and clustering tasks?
Unsupervised dimensionality reduction typically involves deleting
input variables with small variance. The results are extremely
coordinate system dependent.
Supervised dimensionality reduction typically involves deleting input
variables with small
interclassdistance to standard deviation ratios.
Consider two parallel cigar-shaped distributions in 2 dimensions:
Unsupervised dim red would eliminate the dimension perpendicular to
the long cigar dimensions. The resulting classification error rate
would be unacceptably low.
Supervised dim red would eliminate the dimension parrallel to the long
cigar dimensions. The resulting classification error rate would be
appreciably higher.
Hope this helps.
Greg.
|
|
|
|
