how to interpret those different performances?
hi,
I was using different operator settings using Boosting and Bagging with WJ48 and Random Forests...
I basically used an optimize parameter gridsearch, inside it X-Validation, inside that MetaCost operator and AdaBoost or Bagging operator with WJ48 or Random Forest operator inside them....
now I get different performances, I use 70% for training, 30% for testing...:
for AdaBoost with MetaCost and WJ48 Decision Tree I get:
Bagging with MetaCost and WJ48:
Bagging with MetaCost and Random Forest:
now which one of them is most representative? should I use 70 / 30 for cross validation? or something like 50/50 ?
In the last one, I get 83,7% accuracy, however, class 4 recall is only 60%, does it mean I should focus more on that (and therefore, this result is not optimal)?
whereas in the first example, recall is all about 75% for class 4 and 3 and >90% for class 1, and precision is all above 80%, but pred.4 is around 78% only,
but in the last performance , precision is all around 85.6% for pred.4 and 86.7% for pred.3 ...
2nd question: Is MetaCost with Boosting necessary? as I understood, there is already an implicit weighting that weights falsely classified examples more than others...
last question: Can I put more than 1 classifier into AdaBoost and Bagging? (e.g Decision Tree and Naive Bayes or SVM)?
Answers
edit: I noticed when I created the post, the last 1/3 of my post where cut off.. it didn't appear in the thread... is this a bug? it already happened to me twice, when I have more than 2 pictures in there...
Fred,
i would say they are all the same. Go for a cross validation and have a look at the std_dev. Then you will propably see that they are comparable in their variances.
Are you sure you want to focus on accuracy?
For the potential bug: Please consult @stevefarr he can help you.
~Martin
Dortmund, Germany
this may be a lame advice, but make sure to use same "local random seed" for the all the validation, if you leave it at default, you may be training and testing against totally different combinations .. would not impact significantly in many cases, but there is always that one edge case
Hi Fred
Sorry you are having this issue. Have not come across it before.
I just successfully posted a message with 4 images on it. The system settings allow users to add up to 1000 images each.So, I guess something else is causing the issue.
1) The permitted image file types are: *.jpg;*.JPG;*.jpeg;*.JPEG;*.gif;*.GIF;*.png;*.PNG,*.pptx
2) The maximum file size is 10000kB
Could either of these cause the issue?
If not, could you point me to the post that isn't working and send the images you want in the post to community@rapidminer..com. I can then try to replicate the post to eliminate the possibility that my rights on the system differ from yours in the attachment of images.
Thanks
Steve Farr
@stevefarr hm thats weird, I used 3 png images with 10 kb... when I posted first, the lower part of my text with the last image was missing...
I used the "back" button on my browser, copied that part out again, edited my post and pasted it in again, happened to me 2nd time, I cannot explain why...
maybe it could also be display error from my browser, at least I couldn't see the last part of my post....
ok thanks, @mschmitz, why do you mention "are you sure to focus on accuracy" ? Should I use another performance metric ? can you clarify on that?
@mschmitz raises a good point, you should'nt solely focus on Accuracy IMHO. I certainly look at it and the Std dev when comparing differnet performance vectors but in cases on binomimal classification, I also look at AUC and even Kappa. I believe Martin has a link to a research paper on the discussion of AUC under the PR curve(?), which is another interesting measure in model evaluation.
@Thomas_Ott
ok, but where can I see std dev in performance operator? there is no field for that..
and I can't use ROC or AUC, because its 3 classes , would have to use Multinomial to binomial classification or something similar...
You will get a std dev from you accuracy if you use X-Val. The PNG's you posted show me that you're not using X-val at all. You will see something like: Accuracy 70% (+/- 5%)
hmm.. thats really weird, because I was using X-Validation operator all the time...
in this screenshot its there...no idea why its not in the others ?!