5
$\begingroup$

Without going into specifics, I'm currently working on a system that involves 20-25 questions being answered as either Green, Yellow, Orange or Red. After completing a subset of these questions (many questions can be left as defaulting to Green), the system allows our users to choose one outcome out of four, roughly corresponding to the answers they entered (OutcomeGreen, OutcomeYellow, OutcomeOrange or OutcomeRed). The answer that was selected most tends to be a good indicator as to what outcome they will select, but that's not always the case.

After having this system in place for the last 2 years, now I've received a request to have the system itself make a recommendation as to which outcome the user should select. Using data already accumulated over this period, I'd like to get some insight as to which questions/answers tend to be most influential for specific outcomes, and possibly give them more weight when determining what to recommend.

My main dilemma is that my last class on statistics was more than 20 years ago, and just looking through the tags here made me feel that I'm out of my depth. With the description I've provided, and the vast knowledge contained within this SE:

  • Is there anything I should be looking into (tools, subset of CrossValidated tags) that would help gain better insight, and where I should look for more information?
  • Is there a quick way to get up-to-speed on what I'm missing?

Background: I'm a developer in many programming languages, and an amateur mathematician (mostly playing around in number theory and linear programming). I'm also a quick learner; I've been learning how to use R in my spare time. I just need some indication as to where I would find info quickly that would help me move forward with this.

Improve this question
$\endgroup$

3 Answers 3

Reset to default
3
$\begingroup$

Actually, this isn't by most definitions a recommender system, and anything you read in the literature about recommender systems might be geared toward solving a similar but slightly different problem (namely, where the input and output space are the same set).

This is, by most conventional definitions, a classification problem, so looking for tags related to Classification might help.

In terms of actual approaches: as Placidia mentioned, CART and random forest methods are quite popular right now. Additionally, a classic method is Logistic Regression, which might be worth checking at (I'm not an R user, but I believe R has an implementation of it provided).

More than that is tough to provide without a knowledge of what your system is doing and what approaches might be valid. I'd recommend refreshing your basic probability and statistics and thinking about the way that your variables might be related, then taking a look at the classification methods included with R (and their respective Wikipedia pages).

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thank you for helping me with the terminology; coming in with an idea on what I need and what I thought was a descriptive term, only to find out that the term is loaded with other preconceptions is something I experience a lot when developing systems for fields in which I'm unfamiliar. $\endgroup$
    – taserian
    May 12, 2014 at 17:50
3
$\begingroup$

You could try CART (tree) classification regression. That would select a decision tree algorithm for the outcomes based on the answers to the questions. As a bi-product, it would indicate which questions are most important in predicting outcome.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thank you for the guidance; I've got a better handle on what I'm capable of, now that I've poked around in R and not made too much of a mess. $\endgroup$
    – taserian
    May 12, 2014 at 17:51
0
$\begingroup$

Before starting any research you should know about good resources and the research groups working in the filed. May be you can have a read of following paper to get some valuable resources and authors.

"A Scientometric Analysis of Research in Recommender Systems" . Journal of Scientometric Research: 71–84. doi:10.5530/jscires.5.1.1 [PDF]

Improve this answer
$\endgroup$

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.