CS 539: Bayesian Network Structure Learning

Class Times: Tuesdays and Thursdays, 2:00pm - 3:20pm
Class Location: Nash 206
Instructor: Weng-Keen Wong
Email: my last name at eecs dot oregonstate dot edu
Office: KEC 2075
Office Hours: By appointment or open door

Announcements

• (Feb 14, 2006) Assignment #3 is out.
• (Feb 7, 2006) Set u = 4 or 5 for the Alarm network.
• (Feb 7, 2006) If you are getting an AUTON lab message complaining about exceeding some big number of mallocs, compile with the define flag AMFAST in Windows(in Linux, compile with gmake t=fast).
• (Feb 6, 2006) Assignment #2 is due Thurs Feb 9 at the start of class
• (Feb 4, 2006) In Line 10 of the K2 algorithm, there is a term Pred(x_i) which means the predecessors of node xi in the ordering provided to K2. So for example, if you have an ordering A, B, C, D, E, F

  Pred(D) is the set {A, B, C}

• (Feb 4, 2006) Use am_srand(int seed) in utils/ambs.h to seed the random number generator for probabilistic_shuffle_ivec(). If you don't set the random seed, you will get the same random number sequence every time you call probabilistic_shuffle_ivec().
• (Feb 2, 2006)
  1. The mk_empty_bayesnet_from_datset function in netica_auton.c doesn't quite work because you can't get a nodelist from the empty Bayesian network. I've modified a function to return a nodelist.
  2. There is a bug in mk_empty_datset_from_bayesnet in netica_auton.c which was causing s to appear in a w_matrix. I've fixed this as well.

     The fixed netica_auton.c and netica_auton.h files are now in the assn2 download (if you haven't downloaded it yet).

  3. Calling the factorial function on large numbers is numerically unstable. Use the am_lgamma function in utils/stats.h. The am_lgamma function computes the log of the Gamma(x) function. Recall from class that Gamma(x) = (x-1)!. This means you need to do your k2 score function calculation in log space.
  4. When you pass in the in_attnums parameter to mk_ctab_for_w_matrix (when creating an ADTree), you need to make sure the in_attnums ivec is sorted. You can do this with a call to in_attumns = mk_sivec_from_ivec(unsorted ivec) in utils/amdmex.h. This function makes a sorted ivec from a regular ivec. I've updated the autonlib.txt tutorial file to include this comment.

     NOTE: This is very important....

     Because you are sorting the ivec, this messes up the order of the indices of the parent nodes that Netica is expecting. When you call SetNodeProbs_bn(node,parent_states,probs), the parent_states array expects the parents to be in the same order that GetNodeParents_bn returns. In general, this will be different from the order of parents that is in in_attnums (which is passed in to the ADTree functions and consequently, which is the order of parents returned in the values ivec_array).

• (Jan 26, 2006) Typos in notes on Scoring Metrics I. These have all been fixed on the notes off the webpage:
  • pg 8 - first line: P(Mm, D) should be P(Mm | D)
  • pg 9 - first line: P(Mm, D) should be P(Mm | D)
  • pg 38 - first line: P(Bs | D) should be P(Bs, D)
  • pg 40 - first line: P(Bs | D) should be P(Bs, D)
  • pg 41 - last line: P(Bs | D) should be P(Bs, D)
  • pg 58 - The fraction should be (Nijk + 1)/(Nij + ri) where ri is the number of possible values for node i. This needs to be fixed on the programming assignment document as well.
• (Jan 26, 2006) Assignment #2 is out, due Feb 7, 2005 at the start of class. The webpage has been redesigned and the assignments are now in a new section.
• (Jan 20, 2006) Windows port of Assignment #1 source code has been done. Note that Assignment #1 source for Unix has problems compiling on SunOS. See the lecture schedule below for the links to the code.
• (Jan 19, 2006) Assignment #1 code tar file has been fixed. Remember it is due Jan 24, 2006.
• (Jan 10, 2006) Welcome to the course website!

Course Pages

• Course Information
• Lecture Schedule
• Assignments
• Seminar Information
• Project Information