Project presentations (item #1 below) will occur in class on Tuesday, March 9. The project write-up (#2) is due Friday, March 19 at 11:59 PM, however they will be accepted and graded promptly if received earlier. Submit via e-mail to BOTH jiangxu2011 at u.northwestern.edu and ddowney at eecs.northwestern.edu. Use EECS 395/495 Homework 5 as the e-mail subject line. PDF format preferred, though Plaintext, Word, and HTML are also acceptable.

1. (5 points) Prepare a 6-minute project presentation. Send your ppt or pdf slides to ddowney at eecs.northwestern.edu by 10AM Tuesday, March 9. Your presentation should talk about: what your task is, why it's important, and what kinds of results you obtained. Ideally, summarize how your preliminary experiments led you to alter your approach (see below), and whether this led to performance improvements. Also, give one or two suggestions about what you would try next.

2. (15 points) Write up the final results of your project. This write-up should be no more than two pages of text, but can include an arbitrary number of charts and graphs. Briefly explain why your task is interesting, what your original Bayes Net looked like, and whether you feel Bayesian Networks were a good fit for your task. Also describe the experimental results outlined below, along with any other results you found interesting. Give one or two suggestions about what you would try next.

   Note: The most important goal for this assignment is to use your initial experimental findings to devise changes to your original Bayes Net method. The changes may or may not improve performance, but the reasons for trying them should be tenable and explained very clearly.

   Some possible avenues for improvements are: changing your Bayes Net structure significantly, utilizing EM with hidden variables (e.g. by finding a package other than Weka or adapting homework #4 code), obtaining new feature values, switching to an undirected model, or employing a particularly useful prior. You might decide that because your application relies on continuous variables, you'd like to compare Weka's discretization against a different kind of classifier that "more naturally" handles continuous parameters. You might decide to change your task definition to something you believe you can solve, if your initial task was too hard.

   Measure whatever performance metrics you deem most appropriate, training your models on your training set and testing on the test data you set aside at the start of the quarter. Compare the following methods:

   1. Your manually-constructed Bayes Net structure with learned parameters (from homework #3)
   2. Your Bayes Net with learned structure and learned parameters -- note, structure learning is easy to perform with Weka, the tutorial from homework #3 includes this early on. How, if at all, does the learned structure differ from your manually designed one?
   3. Some improved method based on earlier experiments (see discussion above)
   4. Extra credit (1 point): Compare with your manually-constructed Bayes Net parameters (from homework #2)