3
$\begingroup$

I work at a University and have a project to be able to score applicants based on their likelihood to enroll (convert), using their answers to application form questions.

The applications contain name, DOB, date & time of application, country, gender, course selected, English proficiency, funding availability and some other similar fields. There are also various free-text fields, but I think these will complicate things too much to begin with.

My initial thought is to use a regression model to do this, using R. But I am a complete noob - I studied regression at uni 10 years ago....

I have had a search around and I think once I know I'm on the right path I will be able to figure the process out but I am unsure where to begin, and do not want to start by going down the wrong avenue. My main concerns are:

  • Is a regression model the correct approach? If not what is?
  • Are categorical fields a problem - as opposed to continuous fields?
  • There is additional information which is only available for some applicants - can this be included, or do we need to use the same information for all applicants?
$\endgroup$
0

3 Answers 3

2
$\begingroup$

Your question could be closed for too broad, but let's give a try. You want the enrolment probability, this sounds like a logistic regression for me. Neither categorical nor continuous data type should present a problem. You may model the additional information, applications who don't have one simply be assigned a NA category. You can include the NA category in your model.

I recommend you read the book Applied Predictive Modelling, I think it has a section on credit-card applications, which is close to what you are doing. You should try to learn from it.

$\endgroup$
1
  • $\begingroup$ Thanks, I did think it may be too broad, but it is the broader details I'm struggling with. I will check the book out. $\endgroup$
    – agwah
    Jan 22, 2017 at 12:43
0
$\begingroup$

Let's see (broad strokes ahead)

There are two ways of doing it supervised learning and unsupervised learning.

For supervised learning we either need previous years data (with the final enrollments) or you need to calculate the probability of some of the applications(Enough so that the model can predict accurately for new data). Then it becomes a problem of linear regression(as probability will be a continuous variable). If you decide a certain cutoff(threshold) that probability(y =0 below and y =1 above ) then it becomes a problem of logistic regression. Once the data is ready machine learning can be done in half an hour. Now to answer your questions.

  • Yes, Regression model is a correct approach but it's a broad term.
  • Not a problem, you need to make the categorical fields factor using as.factor() in R.
  • you should use same information for all the applicants but that doesn't mean you can't use that info. If the fact that information is available has a bearing on the outcome then make a variable and put it 1 or 0 accordingly. If what's available in that information has a bearing then it gets a little tricky, may be count the number of common relevant words and put that number in a variable.
$\endgroup$
0
0
$\begingroup$

The problem you're trying to solve is called binary classification. There are many algorithms and there's ton of tutorials on it. Actually it is so well studied that ready-made tools are available that try all algorithms and choose the best for your task. https://github.com/paypal/autosklearn-zeroconf is one of them that makes auto-sklearn binary classifier able to run on arbitrary tabular data (like what you have).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.