2
$\begingroup$

Each training example has 100 numeric attributes plus one output class and about 80% of the attributes are 'zero' (means no data collected). The value of attributes varies in a small range, like $(-20,20)$. I have 100 examples like this. What method/classifier should I use? I tried KNN, Naive Bayes, SVM, random forest/tree, none of these methods give me accuracy above 50% (I used 10-fold cross validation). What should I do?

$\endgroup$
  • $\begingroup$ Did you tune the parameters on the SVM? How many trees are you using in your forest? Please add more details regarding the implementation of the algorithms. $\endgroup$ – Firebug Aug 22 '16 at 20:27
  • $\begingroup$ I tried 100 trees,got 44% accuracy,10 trees, got 33% and 200 trees got 40%. For SVM what specific parameter are you looking for? $\endgroup$ – tikael Aug 24 '16 at 15:21
2
$\begingroup$

This is the well known large $p$ small $n$ problem, often named as $p \gg n$, common in biological, biomedical and imaging problems, basically any field where either data is scarce/expensive to come by or simply carries too much information.

There are several techniques often used in this scenario, such as regularization and attribute bagging. Implicitly, you already tested both, respectively on SVMs and Random Forests. You can try to improve upon that, though. Try other regularization penalties, such as elastic net, try using it fused with attribute bagging.

Another possibility is that your data simply doesn't explain your outcomes.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.