EXPLAIN IMPUTATION PLAN

probcomp / bayeslite

BayesDB on SQLite. A Bayesian database table for querying the probable implications of data as easily as SQL databases query the data itself.

http://probcomp.csail.mit.edu/software/bayesdb

Apache License 2.0

917 stars 64 forks source link

EXPLAIN IMPUTATION PLAN #181

Open riastradh-probcomp opened 8 years ago

riastradh-probcomp commented 8 years ago

From @axch: Crosscat should be able to explain why it made an imputation choice it did, phrased as:

What are the columns highly dependent on this column (i.e., in the same view)?
What are the rows that are similar to this row with respect to those columns (i.e., in the same cluster)?
What are the sufficient statistics of the column over the similar rows (i.e., cluster suffstats)?

This can be done exactly for a single model; explaining an integral over all models might be a little more difficult to visualize. No amputations, please.

gregory-marton commented 8 years ago

This echoes #77, and I'd love to have this soon. It might be a candidate for implementation in recipes because those questions are answerable generically, not only in a crosscat-specific way.

Also, should this particular version of the bug be on crosscat's tracker? If so, perhaps the issue here is to plan an interface by which particular models can tell bayeslite about these imputation choices?

raxraxraxraxrax commented 8 years ago

Because of #194 , we might want to accelerate this or something like it into a polish release if it's possible.