Reimplement EMCOMP - Githubissues

morganjwilliams / pyrolite

A set of tools for getting the most from your geochemical data.

https://pyrolite.readthedocs.io

Other

132 stars 36 forks source link

Reimplement EMCOMP #6

Open morganjwilliams opened 6 years ago

morganjwilliams commented 6 years ago

Imputation using the expectation maximisation algorithm EMCOMP includes the assumption that data is below a certain threshold, which must be used as input. Consider reimplementing an expectation maximisation imputation algorithm in light of:

data is not missing at random, but is also typically not missing only due to being below detection (the expectation of the values is not necessarily 'below a threshold')
'detection limits' are commonly not calculated, and rarely available as metatdata

morganjwilliams commented 5 years ago

Working version of EMCOMP is now included in pyrolite (2c04ec0ec3562e042422b633aedc3eb12d3dd79e, see docs).

morganjwilliams commented 5 years ago

The EMCOMP routine is currently unstable (trending towards inf, resulting in nonfinite covariance matrices).

morganjwilliams commented 5 years ago

The current version of the algorithm is now stable. Potential enhancements may follow.