Re: classifiers learning only from one class (should be: Cost-Sensitive learning)

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

Re: classifiers learning only from one class (should be: Cost-Sensitive learning)

Dr.techn. Alexander K. Seewald
Hi,

Over- and Undersampling does not tend to work well. I'd suggest
CostSensitiveClassifier (which tweaks the threshold for the output
probabilities without changing the learner) and MetaCost (which
changes the training set) - both see meta. Both output a single model.

A combination of CSC with NaiveBayes, or MC with JRip or J48 works
reasonably well for a wide range of problems.

Best,
  Alex
--
Dr.techn. Alexander K. Seewald

Austrian Research Institute     [hidden email]
for Artificial Intelligence         www.oefai.at
------------------------------------------------
         Information wants to be free;
Information also wants to be expensive (S.Brant)
--------------- alex.seewald.at ----------------

_______________________________________________
Wekalist mailing list
[hidden email]
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
Reply | Threaded
Open this post in threaded view
|

Re: classifiers learning only from one class (should be: Cost-Sensitive learning)

Paul m
There is evidence to suggest that "intelligent" under/over sampling
(along with base classifier, eg c4.5) outperforms base classifiers,
boost+classifiers & cost sensitive learners.

It is certainly an area that is ripe for further investigation and
some advancement...

/paul

On 7/7/05, Dr.techn. Alexander K. Seewald <[hidden email]> wrote:

> Hi,
>
> Over- and Undersampling does not tend to work well. I'd suggest
> CostSensitiveClassifier (which tweaks the threshold for the output
> probabilities without changing the learner) and MetaCost (which
> changes the training set) - both see meta. Both output a single model.
>
> A combination of CSC with NaiveBayes, or MC with JRip or J48 works
> reasonably well for a wide range of problems.
>
> Best,
>   Alex
> --
> Dr.techn. Alexander K. Seewald
>
> Austrian Research Institute     [hidden email]
> for Artificial Intelligence         www.oefai.at
> ------------------------------------------------
>          Information wants to be free;
> Information also wants to be expensive (S.Brant)
> --------------- alex.seewald.at ----------------
>
> _______________________________________________
> Wekalist mailing list
> [hidden email]
> https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist
>

_______________________________________________
Wekalist mailing list
[hidden email]
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist