-2
1

I'm working with similar data sample as groupon site. I'm triyng to design best algorithm for such domain. Existing approaches seems will not work. For example item-to-item recommendation will not work due to short life of each deal - model will work only for several days. Can you point me to some articles related to recommendation in similar domains?

Training Data: Anonymous user profile, deals profile and transactions.

Target: Build ranking function for each user - i.e. personalisation of deals to users based on their history transactions

asked Dec 13 '11 at 22:49

yura's gravatar image

yura
1025374854

edited Dec 14 '11 at 12:56

Can you give a bit more detail? What data to you have, and what are you trying to predict?

(Dec 14 '11 at 09:21) Noel Welsh

In his new online lectures, Andrew Ng has a good explanation on how to rank for user preferences. Instead of using vectors of features for each Deal, you could take a different approach and use the users as features

(Dec 15 '11 at 02:07) Leon Palafox ♦

Don't have time for a detailed answer right now, so a few notes. The problem you're facing is called a "warm start". That is, you want to start making good suggestions before you've collected a lot of data about the items you're offering. To do this you might consider the following:

  • represent your items and customers as some set of features. (e.g. words in the item description, and whatever you have for customers)
  • build a classifier to predict item x customer -> {0,1}
  • use this classifier to initialise your online recommender system.

Hope that helps.

(Dec 15 '11 at 11:23) Noel Welsh

Probably can do a lot worse than reading this:

http://pages.cs.wisc.edu/~beechung/icml11-tutorial/

(Dec 15 '11 at 17:10) Noel Welsh

@Noel thanks those resources cover topic very well

(Dec 16 '11 at 12:39) yura

I don't understand the negative votes on this question - while the title may see overly specific application focused (which deserves downvotes), the actual question is not and is well defined. Anyone have any thoughts?

(Dec 18 '11 at 18:00) Robert Layton
showing 5 of 6 show all
Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.