So what counts as a "novel" classifier?

Question

[spacegeek1212]

Hello all. I'm working on a research project in which I run a binary classifier on a library of small molecules. I'm working with Java-ML and the libsvm package to implement the SVM algorithm and necessary methods to get results. The project needs to be as original as possible, and my ultimate aim is to be able to call my binary classifier "novel." Here's my question: what would make a binary classifier novel?

[Ahill]

I think it would be very surprising if you had algorithmic novelty, given all the techniques you mention are well established, and there is a massive literature on machine learning. On the other hand, if your classifier correctly predicted a useful class membership for the small molecules, for classes that have never been successfully predicted before, or predicted classes that formerly required more difficult-to-collect molecular features, and this was verified through application to an independent test set, you would have a novel classifier, even if not using a novel classification method.

[Khader Shameer]

You have been asking a bunch of similar questions. Some of us have provided very relevant answers, but however you are keep asking more questions without validation any of your previous questions. I would highly recommend you to spend some time to think through your problem, tie your questions together and ask a well-posed question.

[spacegeek1212]

The questions are similar so I can see how it's getting annoying. However, this is my last question. I'm a high school senior who has had no experience with support vector machines, bioinformatics, etc. so please excuse the redundancy. I can't delete posts, but I would. I don't have a mentor to ask really obvious questions to—this website is all I have. I'm gonna spend some time putting the pieces of the answers I've received together. Sorry for any inconvenience caused.

Answer 1

[Khader Shameer]

"Novel" is a relative term. You can bring in novelty to a machine learning based classifier by

• Application of a machine learning algorithm to a novel
• Incorporating new mathematical/statistical estimates to do classification task
• Apply different existing method / combination of methods that not been implemented to your classification task
• Use new set of feature or hybrid or derived features (For example see our papers where we have combined multiple features). See sections on features: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2901629/ http://www.ncbi.nlm.nih.gov/pubmed/21592079