NEWS.md

In orthoMTL: Multi-Task Learning with Orthogonal Constraints

orthoMTL 0.1.0 (2026-07-10)

Renamed from orthopen to orthoMTL

This release represents a major refactor and scope expansion of the orthopen package (https://github.com/kevinVervier/orthopen).

New features

• Survival analysis mode: survival and censored.mat arguments in orthoMTL() enable censored time-to-event data.
• Elastic-net sparsity: alpha mixing parameter in [0, 1] blending the orthogonality penalty (alpha = 0) with L1/Lasso (alpha = 1).
• Survival data utilities: create_longitudinal_labels(), create_indicator_matrix(), create_constraint_matrix().
• Cross-validation: cv_orthoMTL() with parallel grid search over lambda, step size, diagonal value, and elastic-net mixing.
• Bootstrap inference: bootstrap_orthoMTL() compares real coefficient variability against null-model permutations.
• Evaluation: cindex_mtl() for multi-task concordance index.
• Visualisation suite: plot_heatmap(), plot_correlation(), plot_prediction(), plot_bootstrap().

Breaking changes from orthopen

• Function renamed: orthopen() → orthoMTL().
• Return list key renamed: $W → $B.
• Default disjoint changed from TRUE to FALSE.
• Elastic-net is controlled by a single alpha mixing parameter in [0, 1] (replacing the enet/lambda1 pair). The penalty is lambda * [(1-alpha)/2 * Omega_K(W)^2 + alpha * ||W||_1], ported from orthopen v1.1.0. This fixes the previous inconsistent mixing (L2 ≈ 0.25lambda vs L1 = 0.5lambda1).
• Iso package dependency removed; replaced by internal nnmaxheap_C().

Bug fixes

• Variable T (masking base::TRUE) renamed to numTasks.
• Input validation added for X, Y, lambda, alpha.
• Warning emitted when max_iter is reached without convergence.
• Logistic gradient corrected for labels in {-1, +1} (ported from orthopen v1.1.0): the gradient previously used the {0, 1} cross-entropy form, which is wrong for y = -1.
• Logistic loss made overflow-safe via the stable softplus identity max(-z,0) + log1p(exp(-|z|)) and exponent clipping in the gradient.

Infrastructure

• Package scaffolding: LICENSE (GPL-3), testthat runner, inst/CITATION, GitHub Actions CI.
• Selective NAMESPACE exports.

General-purpose regression & classification support

The solver already supported regression (logistic = FALSE, survival = FALSE) and classification (logistic = TRUE); these are now first-class end-to-end with matching evaluation, tuning, and simulation scaffolding (previously survival-only):

• Non-survival metrics: rmse_mtl(), r2_mtl() (regression) and accuracy_mtl(), auc_mtl() (classification), complementing the survival cindex_mtl() (unchanged).
• Mode-aware cross-validation: cv_orthoMTL() gains logistic and metric arguments. The scoring metric now defaults by mode — cindex for survival, auc for logistic, rmse otherwise — and selection honours each metric's optimisation direction.
• predict() type argument: "link" (default, unchanged), "response" (sigmoid probabilities for logistic fits), and "class" (predicted {-1, +1} labels for logistic fits).
• Simulation modes: simulate_mtl(mode = ...) now generates "regression" and "classification" responses in addition to "survival" (the default).

Solver

• Gradient-step schedule exposed (schedule argument to orthoMTL()). The previously hardcoded sqrt(i) decay is now schedule = "sqrt" (the default, reproducing prior results exactly); "log", "const", and "linear" are also available. An A/B study (validation/ab_s02_gradient_schedule.R) found "log"/"const" reach the same optimum ~12--17x faster than "sqrt", while "linear" can stall. Changing the schedule changes the optimisation path and the exact coefficients, so the default is unchanged. cv_orthoMTL() and bootstrap_orthoMTL() gained a matching schedule argument (passed to every fit), so grid search and bootstrap can use the faster schedules too.

Behaviour change

• cv_orthoMTL(survival = FALSE) previously scored every fold with the survival C-index regardless of the data; it now defaults to RMSE for plain regression (and AUC when logistic = TRUE). Pass metric = "cindex" to restore the old scoring.

orthopen 0.0.1 (2017-01-31)

• Initial release on GitHub.
• Core solver for regression and classification with orthogonal column constraints and disjoint support constraints.
• Two vignettes reproducing Figures 2 and 3 from Vervier et al. (2014, ECML-PKDD).

orthoMTL documentation built on Aug. 23, 2026, 5:10 p.m.