0
$\begingroup$

I have a training dataset for classification problem $X \rightarrow y$. Where $X$ is an $n$th dimensional real vector, $y$ is an integer number in $\{0, 1\}$.

I want to solve the next problem: predict as many lines where $y = 1$ as possible with error level no more than $\alpha = 10\%$. In other words, I want to split dataset ($N$ samples) in two sets: 1) $k$ - samples predicted to be $y = 1$, where $y = 0$ cases less than $k * \alpha$ 2) $N-k$ samples - I don't care about. I need to maximize $k$ with given $\alpha$ .

I have a lot of experience in classification with ML, but this problem is rather unusual to me. Please help me how to formalize this problem for ML. Which and how should I use ML algorithm for the problem? I will be happy to get advice about general idea.

$\endgroup$
0
$\begingroup$

Whether you use logistic regression or discriminant analysis you will get some sort of score for the likelihood that Y = 1. So

  1. Split your data into training/test (this assumes a relatively large sample)
  2. Apply the classification algorithm of your choice on the training data, as long as it gives a score of some sort.
  3. Look at $k*\alpha$ on the test data for various scores.
  4. Pick the one you like best
$\endgroup$
  • $\begingroup$ This the first thing I thought. But I hoped there is more effective way, especially for this goal. Thank you. $\endgroup$ – alex_io Apr 16 '14 at 23:06

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.