new analysis (Migrated from community.research.microsoft.com) RRS feed

  • Question

  • nTH posted on 02-21-2011 3:24 PM

    Can anyone help with the following:


    I have one dataset that contains three values for 10 people.  Based on the inputs: toyota, honda, Chevy, I want to infer that their next car purchase will be possibly a toyota.  So I would train it with the above scenarios giving a final value of which make they will buy.  So if someone buys: honda, chevy, Kia, then they may purchase a Kia having 80% probability (based off the training data). 

    Friday, June 3, 2011 6:19 PM


  • John Guiver replied on 02-22-2011 4:22 AM

    First you need to define what your model is. Your model encodes your assumptions.

    In general we cannot advise about models, only about implementations. However, you could, for example, consider an HMM and learn the transition and emission probabilities. However there will be N(N-1) transition parameters and N(M-1) emission parameters where N is the number of states, and M is the size of the discrete observation space (i.e. the number of car makes in your case). This is a lot of parameters for a small amount of data (10 people) so the answers (the parameter posteriors and the purchase predictions) will be quite uncertain.


    Friday, June 3, 2011 6:19 PM