4

So, I have found an application which could benefit from a machine learning algorithm:

There is a pretty standard method in my area of research for design of experiments and I have discovered that one step could be automated to save time. The automation can be done with a variety of standard machine learning algorithms and I've selected one that:

  1. is used pretty often, according to literature
  2. provides very satisfactory results (high accuracy) on my data

I am currently writing a paper on my findings. My question is, how could I rule out the alternatives or at least justify my choice (for example, in the "related work" section) against other algorithms that could possibly work better/be more suitable?

Considering machine learning is not my strongest skill, I have tried to implement most of the other state-of-art algorithms to compare to, but for some of them I just don't have the knowledge to do so.

*My question is similar to this one: What's the best way to justify your choice of baseline methods for academic paper?, but the difference is that my paper's novelty is about the new application area, rather than a new algorithm.

Improve this question
3
  • In my opinion, how much of your time you want to devote to exploration of the methods depends on how you position your contribution (in any case it is usefull to check the journal where you want to submit your research to see how proposition of new methods is done in your application domain) Although I dont think that the choise of a specific machine learning method could be theoretically justified, because application of machine learning is more an art, there are many different factors that can contribute to the choise of a particular method and its parameters.
    – avenir
    Apr 21 '17 at 0:50
  • What is your research domain (e.g. Computer science) ? The best advice is very dependent on the expectations of publications in your particular domain.
    – Tripartio
    Apr 23 '17 at 11:59
  • For this specific scenario, I am working with manufacturing time-series data and am trying to do some kind of automated partitioning of a signal. So there are many many ways to do it and I have found one that works satisfactory well. How do I proceed without having to compare with every single alternative? I can devote some time to exploring alternatives, but I do not have the necessary background to go into much depth (which means it is difficult for me to implement algorithms from papers with lot of maths).
    – DimP
    Apr 24 '17 at 17:02
2

There is a pretty standard method in my area of research for design of experiments and I have discovered that one step could be automated to save time. The automation can be done with a variety of standard machine learning algorithms and I've selected one that:

It sounds like you are working out a problem in your field that has never before been addressed using machine learning algorithms, and you are the first to try this approach.

is used pretty often, according to literature provides very satisfactory results (high accuracy) on my data

The application of such machine learning algorithm has provided you good results, where I assume that good means at least good enough as the state-of-the-art in the field.

My question is, how could I rule out the alternatives or at least justify my choice (for example, in the "related work" section) against other algorithms that could possibly work better/be more suitable?

If my assumptions above are correct, then your paper's main contribution is to solve a known problem in a novel way. This is sufficiently good an achievement that does not need to be justified. It sounds like your only problem is to convince a community of academics that are not normally exposed to machine learning, that your choice is reasonable (NOT optimal) and your results robust.

In machine learning - both in academia and in industry - you often find yourself using a model that is good enough: it does not have to be the best, nor the most complex; what has to do is do the job correctly and provide interesting results. All you have to do in your paper is then to explain the algorithm and show why the results are robust.

Improve this answer
2
  • 1
    After getting a bit more experience in the field specifically and machine learning more generally since I posted the questions, I personally agree with your answer and the conclusion about being good enough. Also, I have now long submitted the paper following this approach in the end i.e. trying to prove why the results are beneficial, reliable and the research methodology before reaching to this selection of algorithm. I will accept your answer as I think it would have helped me a lot if it was posted a year ago! :)
    – DimP
    Sep 22 '18 at 23:28
  • @DimP thanks, unfortunately I replied as soon as I saw this, knowing it was too late for you, but hoping it would help others. I have been in a similar situation years earlier and had to fight hard to have such a paper accepted - but it was a fight beneficial for the field. Glad to hear you have been successful with it!
    – famargar
    Sep 23 '18 at 20:32

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.