cicdw / dask-glm Goto Github PK

Ability to quickly refit on new variables (with same output), and do model comparisons automatically.

Allow for regularization; can begin with l2 regularization to keep the optimization simple.

Get backtracking line search to work without introducing inefficiencies.

Docstring everything

It would be nice to be able to feed in formulas on top of dask DataFrames similar to statsmodels; related: patsy issue #93

Ensure only methods needed for inputs X,y are: .dot() and .T.

Produce statistical output - start with p-values.

Update the requirements / environment file to reflect any (newly added) dependencies.

Write tests for all the algorithms / computations.

Implement BFGS method and compare.

Allow for Dask DataFrame inputs to track variables by name.

Using the Optimizer as a super class, write a LinearModel / NormalModel class for linear regression fits.

Fix all the confusing sign issues with the algorithms / gradient / hessians.