08.02.2013 Views

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

New Statistical Algorithms for the Analysis of Mass - FU Berlin, FB MI ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.8. EXTRACTING FINGERPRINTS 57<br />

Figure 3.8.18: This shows <strong>the</strong> feature selection (FS, left) and dimensionality reduction<br />

(DR, right) process. First, in <strong>the</strong> FS stage, from a set <strong>of</strong> potential features<br />

(pairs <strong>of</strong> masterpeaks, top image) <strong>the</strong> best x are selected (x = 8 in this example).<br />

Thus, each spectrum is projected onto a point in R x . These points are <strong>the</strong>n projected<br />

during <strong>the</strong> DR stage into R 2<br />

where most clustering algorithms are easily applicable.<br />

� Enhancing generalization<br />

� Speeding up <strong>the</strong> learning process<br />

� Enabling <strong>the</strong> interpretation <strong>of</strong> <strong>the</strong> model<br />

The core idea is to apply a mapping <strong>of</strong> <strong>the</strong> high-dimensional data space<br />

into a space <strong>of</strong> fewer dimensions. Two widespread strategies are filtering (e.g.<br />

in<strong>for</strong>mation gain) and wrapping (e.g. genetic algorithm) approaches. In classification<br />

applications, feature selection may be viewed as a discrimination<br />

problem where one aims to emphasize <strong>the</strong> class separability by a judicious<br />

choice <strong>of</strong> feature parameters. Ideally, <strong>the</strong> extracted features should reveal<br />

some unique non-redundant characteristics that are most effective in discriminating<br />

between classes.<br />

The selection <strong>of</strong> features is necessary because (a) it is mostly computationally<br />

infeasible to use all available features in <strong>the</strong> subsequent machine learning<br />

algorithms and (b) problems emerge when limited data samples but a large<br />

number <strong>of</strong> features are present (this relates to <strong>the</strong> so called curse <strong>of</strong> dimensionality).<br />

It can be shown that optimal feature selection requires an exhaustive

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!