Hi, my name is Matt S Rinc and can help you - programming for 20 years, webmastering and now more and more engaging in statistics (three years of R scripts, now also PhD student of Knowledge Management with ... statistics).
It's hard to predict exact number of votes because for completely random user choices you would need 100 votes to predict next at 95% accuracy. But usually users have some preferences and that number drops quickly - based also on how many categories you have.
So, I would like to see your data. What you need is creating a model with the test data (to learn the users' behavior), then run it with the sample data - all of course parts of data you already have. Then using different sets of test and sample data find best model to recommend for next user's vote prediction.
After that there would be few formulas based on best fit categories that a user currently false into and that could be scripted. These would take place after the best say 80% or 90% or 95% (most commonly used in statistics) accuracy is achieved per determined model, say to be 15 votes etc.
OK, I do develop one Mac application to finish next week but could help you here. But I would like first to see at least part of the data (say 20 or more series of user votes and categories of coupons - and how you want the coupons to be segmented, giving you most benefit).
Looking forward to your reply,
Matt