search

kearnz / autoimpute

Python package for Imputation Methods
MIT License
241 stars 19 forks source link

Priors for Bayesian Methods #18

Closed kearnz closed 5 years ago

kearnz commented 5 years ago

Default value of priors for mu and sd in numerical bayesian methods for imputation should likely be from the observed series for which we are trying to predict. Right now they are hard-coded at 0 and 10 respectively. Eventually, the user will have the option to pass these values, but until then, we still likely need better defaults.

Bad priors can lead to slow convergence for MCMC and even numerical overflow issues with pymc3 when realistic values are not within the prior interval.

kearnz commented 5 years ago

Imputation methods now take parameters for priors as arguments to respective function. That being said, strategy_args not supported yet, so PredictiveImputer still used default values. That being said pmm now sets prior mu intelligently using point estimates from least squares.

kearnz commented 5 years ago

PredictiveImputer (and MultipleImputer) take imp_kwgs, so parameters can be set more intelligently. defaults still in place if the user does not pass any imp_kwgs.