New Behavior

• test and support python 3.8, 3.9, 3.10, 3.11, and 3.12

Bug Fixes

• use dynamic versioning to automatically set package version. fixes https://github.com/dswah/pyGAM/issues/339 https://github.com/dswah/pyGAM/issues/336

New Features

• move to Poetry for dependency management
• use flake8 + black for linting

Bug Fixes:

• use int instead of np.int

Breaking changes

• remove support for python 2

New Features

• cyclic p-splines: you can now train models with periodic features by using the 'cp' basis like so:

GAM(s(0, basis='cp'))

• factor smooths now allow dummy coding, via:

GAM(f(0, coding='dummy'))

Models using this coding scheme are more statistically interpretable , and computationally less expensive than those using one-hot encodings.

Bug Fixes

• models can mix constrained terms and un-constrained tensor-terms
• tensor terms can be constrained

Bug Fixes

• Fix not None element existance judgement bug in terms.py thanks @BeefOnionDumplings !
• Added a warning issued in summary indicating that there is likely a bug in the p-values

Bug Fixes

• fixed bug where np.int64 did not count as integers. the following no longer fails:

LinearGAM().gridsearch(X, y, n_splines=np.arange(5, 10)).summary()

New Features

• Documentation!!!

Bug Fixes

• removed WIP method randomsearch

New Features

• gridsearch(...) allows searching across a predefined grid of points, without doing the cartesian product, when grid is a np.ndarray of shape (n_points, len(flatten(gam.lam))). This is useful for RandomSearchCV - style behavior.

Bug Fixes

• estimate_r_squared(X, y) no longer raises AttributeError
• dtype=auto no longer allowed for terms
• intercept.lam = None

New Features

• easier global arguments for terms

GAM(s(0) + s(1), n_splines=10).fit(X, y)

will broadcast n_splines=10 to all terms

Bug Fixes

• fixed inconsistencies in GAM instatiation, where

GAM(lam=0.6).gridsearch(X, y)

worked for multi-dimensional X

but not

GAM(lam=0.6).gridsearch(X, y)

New Features

• tensor product terms and feature interactions. On top of that, construction is more precise and less verbose:

GAM(te(0, s(1, n_splines=5))).fit(X, y)

• the partial_dependence() method can return meshgrids to help you make 3D plots of interaction terms
• ExpectileGAM: for creating a non-parametric description of a distribution. Instead of just modeling the mean of a response, we can model any quantile using

ExpectileGAM().fit_quantile(X, y, quantile=0.25) 

Breaking Changes

• GAM construction is different but much simpler. check out the docstrings for help.
• generate_X_grid and partial_dependence methods require you to specify term= instead of ~feature=~

New Featrues

• all GAM classes have a verbose argument. this makes them compatible with sklearn GridsearchCV + RandomizedsearchCV
• add toy_classification dataset
• move generate_X_grid to GAM method

Bug Fixes

• users should get a more pythonic experience with partial_dependence by never needing to index with i+1
• _initial_estimate() method no longer fails on value nudge for purely integer observations
• regenerate images
• bugs in readme
• fixes bug where poorly conditioned matrix would fail when using skcholmod
• make2d should not be verbose in initial_estimate()