Attribute discovery - how?

Stefan_EStefan_E Member Posts: 53 Maven
Hi all,

I've come across the following problem:

A data set consists of examples with coordinates in x/y (x, y being integers) and contains labels A/B/C. Problem is to predict which examples are labeled B.

Visually, I discovered that individuals with labels B are immediate neighbors of labels C: (x+1/y = C ==> x/y = B). So, had I defined attributes like

  label_at_(x+1)_y, label_at_(x-1)_y, label_at_x_(y+1), ...

it would have been quite easy for eg. a tree classifier to figure out which individuals had label B, given knowledge on A and C labeled examples.

However, domain knowledge does not easily imply such a neighborhood relation. Hence the question: Are there automatic ways (inside/outside RM) to discover such derieved attributes automatically?

Where do I start looking? Any pointer welcome!

Kind regards                    Stefan


  • Options
    IngoRMIngoRM Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, Community Manager, RMResearcher, Member, University Professor Posts: 1,751 RM Founder
    Hello Stefan,

    are you already familiar with the operator "YAGGA" (an acronym for "Yet Another Generating Genetic Algorithm")? This operator was designed for the automatic detection of such attributes depending (non-linearly) on others. Maybe this helps.

  • Options
    Stefan_EStefan_E Member Posts: 53 Maven
    Hi Ingo,

    thanks a lot for the hint! If I understand correctly, YAGGA tries to create new attributes as a non-linear function of existing attributes of the same example - so that will likely not help here...

    Still, the pointer was useful as it puts me in the direction of value series. So, some trials are due... Let's see what the next weekend brings  :)

    One more question though: the YAGGA2 references a paper of you, RobutsGP: Intron-Free Multi-Objective Feature Construction. Is that available anywhere (Goole didn't find it  :'( )

    Greetings - Stefan
Sign In or Register to comment.