You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I still don't have an overview and design for the desired module structure, especially where to put helper functions and model/estimation functions that are reused in several other parts.
We need to avoid circular imports and need some main modules for user access.
robust_linear_model (current)
resistant_linear_model
to include DetS, DetMM (user facing)
needs deterministic starting sets using robust mahalanobis, and start subset OLS
scale (current, including Mscale w. fixed mean)
not clear yet, simultaneous estimation, mean-scale as in scale.Huber, or leave it to regression models
covariance: scatter or covariance (at normal reference distribution)
to include ogk, tyler, m-estimator for cov and mean
to include here or separate: covDetMCD, covDetS
needs starting sets, ogk, ..., and rank and other covariances in stats
exog-outlier weights, Mallows, ..., not clear whether and where to add this (e.g. reuse GLM with weights)
multivariate linear model, eventually (user facing)
based on covariance but with mean model instead of constant means
multivariate stats, pca, ..., not sure, large literature as application of robust cov/scatter
nonlinear model, eventually, old experimental PR (user facing)
GLM, poisson, ... (no idea about module name, no details yet)
regularized, penalized versions of above, include alongside main unregularized models/estimators
...
helper functions and classes
norms current
robust.tools: currently tuning parameter, efficiency, breakdown point for univariate
multivariate, cov tuning parameter ?
api to collect user facing functions and classes
other
more cov_types, e.g. h4 and HC, HAC
prior weights freq_weights and var_weights (one R package maybe MASS rlm has both, one at a time)
...
Currently I'm not clear where to put the "det" starting set helper code
If I want to put covdetmcd and covdetS in covariance, then I have to put all supporting code also in there to avoid circular imports:
Starting sets needs cov/scatter like ogk, ....
We need application specific iter functions to iterate in DetXxx classes.
Then DetXxx classes need to go into user facing module.
I don't know if lazy imports will be able to handle all the circular imports.
The text was updated successfully, but these errors were encountered:
parking an issue here for now: scaling of norm function
joint objective function for mean parameters and scale in analogy to MLE.
tyler uses min rho(u / s) + log(s) or analogue in multivariate cases
If we use M-scale, then the estimator can be derived with an extra scale_bias factor b (AFAIR) min rho(u / s) + b * log(s)
When we use rho only or appropriate b, then the scaling of rho is irrelevant. Currently rho is scale so that max weight = 1.
If we use Tyler objective function, then scaling of rho is relevant.
For example for t-distribution including scale, we would have to rescale the rho function so that it corresponds to the central part of the negloglike (central part that has the term that depends on mean parameters, which is currently rescaled in StudentT norm).
roadmap for parts
#9133
#9148
I still don't have an overview and design for the desired module structure, especially where to put helper functions and model/estimation functions that are reused in several other parts.
We need to avoid circular imports and need some main modules for user access.
robust_linear_model
(current)resistant_linear_model
scale
(current, including Mscale w. fixed mean)scale.Huber
, or leave it to regression modelscovariance
: scatter or covariance (at normal reference distribution)stats
covariance
but with mean model instead of constant meansnorms
currentrobust.tools
: currently tuning parameter, efficiency, breakdown point for univariateapi
to collect user facing functions and classesCurrently I'm not clear where to put the "det" starting set helper code
If I want to put covdetmcd and covdetS in
covariance
, then I have to put all supporting code also in there to avoid circular imports:Starting sets needs cov/scatter like ogk, ....
We need application specific
iter
functions to iterate in DetXxx classes.Then DetXxx classes need to go into user facing module.
I don't know if lazy imports will be able to handle all the circular imports.
The text was updated successfully, but these errors were encountered: