Created a new transfer learner (called SEER) to overcome challenges of HDP/Bellwether that were first highlighted here
SEER uses causal inference learning (derived from the cause-effect paper). See how it works here.
Compared SEER vs HDP (current State-of-the-Art for heterogeneous transfer learners). SEER comprehensively outperforms HDP in every single case. See sample results.
Also, there exist Bellwethers in HDPs too! Not the same as a homogeneous, but bellwethers nonetheless. Note: I'm going to look closer into this, but many cases like ones in issues #4 and #5 seem to indicate this
Compared SEER vs TCA+ (the current best homogeneous). Again, SEER outperforms TCA+ in almost every case. See results.
Doing
SEER vs. Bellwethers. No comments yet on this one. They both seem to do the same.
One more transfer learner to implement CCA.
Ground work for LN this week: (1) Mutate the tags x%; (2) Increase/decrease no. of tags; (3) Comments on "tag quality". At what point do we notice changes if any?
Todo
By Thursday (Oct 27th): Report for LN folks.
Implement discrete-cause-effect test and use that to validate XTREE's plans. <- My very next reasearch item before Christmas Vacations.
Roadblocks
Last Week:
Cause effect was not meant for discrete distributions.
Causality not for Transfer Learning. This is deprecated, my results show otherwise.
This week: No impediments. Found a paper that talks about cause-effect tests for purely/partially discrete distributions. Needs convex optimization. See this issue for more details.
Over the past 2 weeks:
Done
Doing
Todo
Roadblocks
Admin Issues?