6
$\begingroup$

I am working on this project where I have to predict the acceptance probability of a literary article, let's say a research paper by a publisher, which in this case would be a Journal.

I would like to use a machine learning algorithm for this project which would learn from the pool of articles/research papers published by the Publisher/Journal till date. Later I want to use the current article/research paper at hand to predict the probability of acceptance of this article/research paper on the insights that the algorithm has learnt. Here are my queries:-

  1. What algorithm are there to tackle such task?
  2. How should I make these algorithms learn the word vector, grammars, topics etc?
  3. How do I convert the article into a feature set from which the algorithms can learn?
$\endgroup$
4
  • 1
    $\begingroup$ 0. What dataset do you plan to use? $\endgroup$ Jan 4 '16 at 21:42
  • 1
    $\begingroup$ "learn from the pool of articles/research papers published by the Publisher/Journal till date." - if you "learn" from a dataset of papers that have all been accepted (as presumably, all published papers have been) then your machine learning algorithm will "learn" that all papers are accepted, and will always predict 100% probability of acceptance. On the bright side, your algorithm will have 100% accuracy on your training set! $\endgroup$
    – ff524
    Jan 5 '16 at 8:19
  • $\begingroup$ Hi, @FranckDernoncourt what I was trying to do was learn from the patterns of published paper and use that in the present paper at hand to do the prediction. I get what you implying, I should have a training datasets consisting of mixture of papers every published and papers rejected so that I can classify the testing datasets to check for accuracy and then predict the current paper at hand. Actually I am not get how to approach this task (weather in form of supervised learning or unsupervised learning). $\endgroup$
    – gourxb
    Jan 6 '16 at 5:23
  • $\begingroup$ But ML is better at determining topic. A journal is going to look for new topics not duplicate. Acceptance is more about quality of the article and the research. Not having articles rejected is a key component. $\endgroup$
    – paparazzo
    Jan 6 '16 at 21:36
4
$\begingroup$

Applied Predictive Modeling book has a case study of acceptance rate of grant proposals, which was a Kaggle competition. You can get many good ideas from there.

  • from the dataset, what kinds of features you would look at
  • from the book, step-by-step explanations of how to build models
  • from the Kaggle competition, approaches by the contestants

I think it is very hard to predict the quality of the paper from the text of the paper, but you might have much more luck by including features such as academic rank of the author(s), their past publishing history and success, acceptance rate of the journals the paper was submitted to, and things such as even the length of the paper. Such proxy features for paper quality actually might be very successful...

$\endgroup$
1
$\begingroup$

First of all I think that what you're trying to do is very difficult. A research paper success depends not only of the words but the math, the time when it was published, the journal, etc. You have many features you should take into account.

I would try deep learning. For input, I would add all the features above and perhaps more. As the output I would use the number of citations or maybe something more sophisticated like a custom function of popularity. It's not the same to be published in Nature than in other journals.

$\endgroup$

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.