NEWS

densemlp 0.9.0

weights argument in densemlp(): non-negative case weights, one per row, so the network minimizes a weighted empirical risk (importance weights under covariate shift, survey or inverse-probability weights). Only relative weights matter. Weights enter the loss and gradient of every task (a weighted Breslow partial likelihood for Cox, a weighted IPCW Brier score with a weighted censoring Kaplan-Meier and time grid), the batch-normalization statistics, the input and outcome standardization, and the validation loss used for early stopping. Unweighted fits are unchanged.
weights argument in densemlp_metrics(): weighted versions of every metric (weighted AUC and C-index count each pair as the product of its weights).
weights argument in cv_densemlp() (folds are fitted with their training weights and scored with weighted metrics, i.e. importance-weighted cross-validation) and tune_densemlp() (candidates ranked by the weighted validation loss).

ncores > 1 (ensembles, cv_densemlp(), tune_densemlp()) could hang when R uses a threaded BLAS such as MKL or OpenBLAS: forked workers deadlocked in the BLAS/OpenMP thread pool inherited from the parent. Each worker now runs its BLAS on one thread (new import: RhpcBLASctl).

groups and group_penalty arguments: every level of one or more grouping factors gets its own learned intercept on the linear predictor, trained jointly with the network (fixed-effect style cluster or period intercepts, as in panel data), with one ridge penalty overall or one per grouping factor. New group_effects() extracts them.
offset argument, in densemlp() and predict(): a fixed term added to the linear predictor (a vector, or a one-column-per-class matrix for multiclass).
task = "count": Poisson log-likelihood with a log link, for non-negative counts.
task = "binary" accepts numeric proportions in [0, 1] (a fractional or quasi-binomial outcome).
validation accepts a vector of (at least two) row indices to hold out, for grouped or time-ordered validation splits.
hidden_units = integer(0) fits a linear model (no hidden layers), useful as a GLM baseline sharing the same offset and group terms.

The binary validation (and reported training) loss now uses the full Bernoulli cross-entropy. It previously dropped the (1 - y) log(1 - p) term, so early stopping for task = "binary" tracked a criterion that rewards over-predicting the positive class. Gradients were already correct.

Add src/Makevars / src/Makevars.win linking the numerical kernels against R's BLAS/LAPACK ($(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS)). Without the explicit link line the Armadillo calls (dgemm_, dgemv_, ddot_, dsyrk_, ...) were left undefined at link time on the CRAN Windows builder. No user-visible changes.

The model backend no longer uses torch. Forward propagation, backpropagation and Adam optimization are now implemented natively in C++ via RcppArmadillo, with hand-derived closed-form gradients. The torch (and transitive libtorch) dependency is gone, so the package installs and trains without downloading a deep-learning runtime. ggplot2 is likewise no longer a dependency; plotting uses base graphics.

task = "survival": train against a survival::Surv(time, event) outcome with either loss = "cox" (batch-wise Breslow-tie Cox partial likelihood, the default) or loss = "brier" (discrete-time hazard head trained on the IPCW integrated Brier score). predict(type = "survival") returns the survival-probability curve for Brier-loss models; densemlp_metrics() reports Harrell's concordance index; densemlp_integrated_brier_score() evaluates a Brier-loss model's IBS on new data.
Formula interface: densemlp(formula = y ~ ., data = df) (including survival::Surv(time, status) ~ .) as an alternative to x/y. It must be passed by name, since x/y keep the first two positional slots.
residual, gated, learned cross-feature interaction, exponential moving-average weights (ema_decay), learning-rate schedules (lr_schedule), internal bootstrap ensembles (ensemble, ncores) and a tune_densemlp() grid search.

New batch_norm (default TRUE): set to FALSE for Linear -> ReLU hidden blocks with no normalization and no BN affine parameters.
New input_projection: an optional bare linear layer mapping the encoded predictors to a chosen dimension before the first hidden block. Cannot be combined with interaction.
task now accepts "binary" and "multiclass" explicitly (in addition to "regression", "survival" and the default "auto").
Removed the torch-era arguments that no longer apply: activation, optimizer, weight_decay and input_projection's old list form. The fitted object's class is now "densemlp" (was "densemlp_fit"), and its autoplot() method is replaced by plot() / plot_history().

perm_importance() (model-agnostic permutation importance, now also covering survival) and plot_history() are kept, on the new backend.

New cv_densemlp(): k-fold cross-validation with the same formula/x-y interface and task-aware defaults as densemlp() / tune_densemlp().
Standardized the maintainer's family-name casing to "El Badisy" in Authors@R.

First CRAN release. No NEWS.md was kept prior to this version; see the git history for the earlier development log.