safareg: Structured Additive Factorization Regression

Add-on pacakge for the R package deepregression to fit structured additive models when the number of factor levels is prohibitively large and needs an efficient implementation, when fitting varying coefficient terms with a factorization approach, factorized effects in general, (higher-order) factorization machines, or additive (higher-order) factorization machines.

Requirements

See those of deepregression.

Usage

In (one of) your formula(s) in deepregression you can use any function (e.g., fm) and pass the respective processor (e.g., hofm_processor) via the additional_processors argument in deepregression. For example:

mod <- deepregression(
  y = y,
  list_of_formulas = list(~ 1 + age + fm(V1, V2, V3, V4, V5, V6)),
  data = data,
  family = "bernoulli", 
  additional_processors = list(fm = hofm_processor)
)

The following processors are available:

• fac_processor: efficient computation of categorical effects with many factor levels
• interaction_processor: same as fac_processor but for interactions of two categorical effects
• vc_processor: efficient computation of varying coefficients (interaction of smooth and one or two categorical effect)
• am_processor: same as vc_processor but using an linear array model-type formulation (only for one categorical effect)
• fz_processor: computes a matrix factorization for two or three categorical effect interactions
• vf_processor: same as vc_processor with two levels, but using a factorization approach as in the fz_processor
• hofm_processor: computes (higher-order) factorization machines for a given set of features
• afm_processor: computes additive factorization machines for a given set of features
• ahofm_processor: computes additive higher-order factorization machines for a given set of features

Citation

When using or referencing the contents of this package, cite

@InProceedings{FaStR,
  title={Factorized Structured Regression for Large-Scale Varying Coefficient Models},
  author={David R{\"u}gamer and Andreas Bender and Simon Wiegrebe and Daniel Racek and Bernd Bischl and Christian M{\"u}ller and Clemens Stachl},
  year={2023},
  publisher={Springer International Publishing},
  booktitle="Machine Learning and Knowledge Discovery in Databases",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="20--35"
}

for the efficient factor effect implementation or (time-varying) factorization approaches and

@InProceedings{AFM,
  title = {Scalable Higher-Order Tensor Product Spline Models},
  author = {David R\"ugamer},
  year={2024},
  booktitle =    {Proceedings of The 27th International Conference on Artificial Intelligence and Statistics},
  series =   {Proceedings of Machine Learning Research},
  publisher =    {PMLR}

}

for (higher-order) factorization machines or additive (higher-order) factorization machines.