[Work Log] FIRE - streamlining; missing data

May 12, 2014
Project FIRE
Subproject Piecewise Linear Clustering (tests)
Working path projects/​fire/​trunk/​src/​piecewise_linear/​test
SVN Revision 16784
Unless otherwise noted, all filesystem paths are relative to the "Working path" named above.

Refactored kjb::Matrix::resize to be re-use allocated space under more circumstances. This is currently a bottleneck in inference, and signficantly improves runtime.

Implemented Missing data handling:

  1. Likelihood ignores missing values
  2. initial model estimation handles gracefully handles missing data

TODO: Test missing data. Randomly convert observations to missing, see if cluster parameters and memberships are still correctly estimated. Is there a "critical point" above which missing data ruins results?

Posted by Kyle Simek
blog comments powered by Disqus