Define imputation (predict_confidence)

[axch]

• Define it probabilistically
• Implement the definition as an (approximate?) algorithm against joint simulation and assessment
• Compare speed and quality against the structural implementation in crosscat

[fsaad]

Related probcomp/crosscat#54

[gregory-marton]

@axch, please give more context?

[axch]

We really do not have an operational probabilistic definition of imputation. I wrote a whole essay about it, currently at https://github.com/probcomp/bdb-experiments/commit/de246884daf3043486805c28a5ed05dacf2ecb18#diff-aeb232468aa592f22b8141343fda8001R913

[fsaad]

An upcoming project of mine is to revisit all these "define" issues, define them, then create generic implementations of BQL using Monte Carlo estimates obtained from the GPM simulate and logpdf.

[gregory-marton]

Note that probcomp/crosscat#54 says

We are going to need to be loud about this.

and I agree. @raxraxraxraxrax , I wonder if we can improve caveats in our messaging around what bayeslite is, to try to expose and explain some of these issues?