0

I devised a classification algorithm which is useful for specific complicated distribution of classes of data. The method works good on the artificial data which cannot be solved by typical algorithms in the literature. But yet i couldn't find any real data to evaluate the algorithm on real problems. Also there is mathematical support for the basis of the approach and the way it works. So i was wondering if an algorithm with good mathematical support but no evaluation on real data has any chance to get accepted in computer science conferences like SDM or CVPR?

  • This seems like a technical question — more about how to appropriately evaluate and test an algorithm than about the conference acceptance process. Might it be better suited to cs.stackexchange.com? – PLL Oct 4 '16 at 9:46
  • In my experience in other related fields it is a rare exception to have anything tested on something resembling real data - so much so that investigations on a method using real world data can be a separately publishable result in and of itself (if the results are interesting). But I imagine this varies tremendously by sub-field, so I am unsure if this is true for your specifically stated venues. – BrianH Oct 4 '16 at 17:08
  • Do you have significant theoretical results (i.e. proofs)? – Bitwise Oct 4 '16 at 18:11
  • yes i do have a mathematical proof for the basis of my approach. – Bob Oct 4 '16 at 19:42
3

But yet i couldn't find any real data to evaluate the algorithm on real problems.

There exist a good site on benchmark datasets in machine learning, uci machine learning repository that you can use these datasets as well as the artificial tested data.

So i was wondering if an algorithm with good mathematical support but no evaluation on real data has any chance to get accepted in computer science conferences like SDM or CVPR?

The answer is yes if your approach is backed by theoretical support. On validation of attained results, you should provide intuition about how your proposed approach works on a class of problems via the theoretical results if you have such ones'.

  • my problem with UCI dataset is that, it is difficult to visualize most of them to see the structure of the classes and distribution of the data as they mostly are high-dimensional datasets. – Bob Oct 4 '16 at 12:32
  • You can use Iris datasets and also see this example to visualize this dataset effectively, sebastianraschka.com/Articles/2014_python_lda.html. Furthermore, for most of the dataset you should have a preprocessing step to reduce the number of features to have an insight of the overlapping structure among the classes. – Hadi Oct 4 '16 at 16:26

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.