CSE454 prac notes 2005

These are general comments; you should be able to identify which, if any, apply to you.

• Do start the pracs early.
• Use correct scientific language and style.

• Prac 1: As discussed in lectures, Snob likely finds more than 3 classes because of correlations in the data:

  Length and width are both ~size and are very likely to be correlated

  (note those rather diagonal scatter plots that some of you drew).

  A crude way to remove some of this effect is to use a transformation such as

  <length, width> ---> <length, aspectRatio> where aspectRatio=width/length, or variations on similar themes.

  E.g. RGB-colour values have a similar "problem" and are often transformed <red,green,blue> ---> <brightness,chr₁,chr₂> for the same reason.

  A table of species v. Snob classes shows that the program is something like 95+% "accurate", which is very good indeed. If a machine learning method gets results half-way close to such a figure on real-world data it is doing very well indeed! And this is a very tough test because we have not even told the program that there are 3 species nor which species a specimen belongs to. The program is playing the botanist given some new, unknown specimens. And it only has four numbers per specimen to go on!

  Bites: Finding just one class does not imply that there is no effect, that the data are random; that does not necessarily follow (it might follow, given more details, but it does not necessarily follow). 'One class' is not the "null hypothesis".

  What is the null hypothesis? (You might get some ideas from a completely [different] problem.)

  E.g. There could be a mean close to the full-moon and a small std. dev. --

  you should state the std. dev. corresponding to kappa (κ), and the angles corresponding to at least full- and new-moons.

  Prac 2: ...

  ..., OK, you'll start this one early?