missoNet: News and Changes

1.5.1 — 2025-09-01

Major release. Architecture rebuilt for speed, stability, and cleaner workflows. Recommend all users upgrade.

Highlights

• 3–5× faster on typical problems via rewriting core kernels in RcppArmadillo.
• Adaptive grid and lambda search that scales to large problems with fewer evaluations.
• Element-wise precise penalty control for both coefficients and network edges.

Algorithm & Features

• Consistent search strategy in missoNet() and cv.missoNet().
• penalty customization through beta.pen.factor and theta.pen.factor, enabling exclusion of predictors or structured sparsity patterns.
• relax.net option for de-biased network estimation: refits the active edge set without L1 penalties to reduce bias in the precision matrix.
• Support multiple graphical lasso backends (e.g., glassoFast (default) / glasso) with automatic selection and fallbacks.
• Smarter defaults and data-aware initialization tuned to dimensionality, missingness, and conditioning.
• Full parallelization for cross-validation and grid evaluation.

Performance & Robustness

• Vectorized algebra and memory reuse; internal benchmarks show 3–5× speedups
• The warm-start strategy has been rewritten to enhance stability.
• Tighter line search and stopping in FISTA; fewer oscillations.
• Epsilon‑safe numerics; strict finite‑value checks.

Visualization & UX

• Upgraded S3 plot() for publication-quality output.
• Clearer progress and diagnostics (stages, backend, convergence).

Documentation

• New vignettes: introduction, advanced features, case studies.
• Cross-references and examples aligned across missoNet() and cv.missoNet(); arguments and return values documented consistently.

Bug Fixes

• Fixed Inf propagation in preprocessing.
• Fixed cross-package reference linking.
• Removed unused external data to reduce package size.
• Miscellaneous bugs.

Compatibility: Minor API changes; refined defaults. See ?missoNet, ?cv.missoNet.

1.2.0 — 2023-07-18

• More stable convergence in difficult edge cases; fewer premature stops.
• Improved initialization with better defaults.
• Removed glmnet dependency.
• Better diagnostics and messages.

1.1.0 — 2023-05-10

• Final evaluations are now based on warm starts.
• Tuned default parameters based on empirical testing across representative datasets.
• Dropped C++11 requirement.

1.0.0 — 2022-10-06

• Initial CRAN release.