glmdisc 0.8.1

Version 0.8.0 was submitted to CRAN but not released; its changes are listed below. This version answers the comments of the CRAN team.

• The documentation of every exported function now has a \value section that describes the class and the meaning of the result (plot() is documented as called for its side effects).
• New examples for print(), summary() and predictlogisticRegression().
• plot() restores the graphical parameters of the user (it used to leave ask set to FALSE).
• The arXiv paper is cited through its DOI in the Description, Authors@R contains the ORCID iDs of three authors and the URL field lists the package website and its repository.
• The plot test no longer leaves an Rplots.pdf in the test directory.

glmdisc 0.8.0

This release re-submits the package to CRAN, from which it was archived on 2023-02-26 because the maintainer’s email address was undeliverable (the maintainer address is updated).

Bug fixes

• The SEM-Gibbs chain now actually iterates. After the first sweep the quantized values were stored as character vectors, nlevels() of which is 0: every feature was declared “disappeared” and glmdisc() stopped at iteration 2, returning a random quantization. On a simulated example (5 continuous features, 3000 observations, true model piecewise constant) this gave a test Gini of 0.62 against 0.75 for a plain logistic regression; the fixed algorithm reaches 0.76.
• The validation log-likelihood used when criterion is "aic" or "bic" with a validation set ignored the observations of class 0.
• The BIC used log(n) instead of log(n_train) as penalty.
• When test = TRUE and validation = TRUE, the best iteration was selected on the test set and the performance reported on the validation set (the sets were swapped). Selection is now done on the validation set and the performance reported on the test set, as documented. proportions is c(test, validation).
• Interactions involving a feature were dropped by substring matching (V1 removed V10:V12), which broke every dataset with 10 features or more.
• normalizedGini() depended on the order of the observations when predictions are tied (always the case for discretized models). It is now computed from ranks (2 AUC - 1).
• predictlogisticRegression() aborted the R session when its arguments were not conformable; it now raises an error.
• predict() on a glmdisc object keeps one output per input row and works through stats::predict() as well (an S3 method is now registered).
• Missing values of factor features are handled (as a level of their own).

Changes

• The core of the algorithm was rewritten on integer-coded quantizations: no more design matrices rebuilt at each iteration, one logistic regression lookup instead of one per level.
• The link between a continuous feature and its quantized version (reg_type = "poly") is fitted by a dedicated penalized Newton solver (C++, RcppEigen) with warm start instead of nnet::multinom: about 9 times faster for the same likelihood. A sweep on 3000 observations and 5 features takes about 25 ms.
• best.disc[[1]] is now a regular glm object fitted on the training and validation sets (summary() gives standard errors); best.disc[[2]] holds the fitted links. The encoder element of parameters was removed.
• Integer columns are accepted as numeric features; character and logical columns as factors.
• The Metropolis-Hastings search of interactions now samples the distribution it targets: the BIC is halved (the thesis used exp(-delta BIC), the square of the BIC posterior), the asymmetry of the proposal is corrected, and the pairwise prior is clamped to [0.01, 0.99] (a prior of exactly 1, which happens for a strong effect such as an XOR, made the corrected chain unable to enter the pair). On the benchmark datasets the performance is unchanged within noise; the rule of the thesis can be restored with options(glmdisc.mh_scale = 1, glmdisc.mh_hastings = FALSE, glmdisc.mh_clamp = 0).
• New optional argument patience to stop the chain when the criterion stops improving.
• fast is deprecated and ignored (sampling is always vectorized).
• contr.ltfr() (a re-export wrapper of caret) was removed.
• Dependencies reduced to gam, MASS, Rcpp, RcppNumerical (caret, dplyr, magrittr, nnet, NPflow and Rfast are no longer needed).
• The plot/predict methods are documented as S4 methods; the CITATION file was modernized.
• CI: standard r-lib/actions workflows (check on Linux, macOS and Windows; coverage; styler; pkgdown).
• Added inst/benchmarks/real_data.R, a comparison with other ways of preparing features for logistic regression on real and simulated data.