Skip to contents

Changelog

Source: NEWS.md

L0glm 0.1.0

Initial version

  • Sparse log-concave NMF (splogconNMF) recovers elution profiles and spectra with local native updates, spectral fusion and OpenMP. Supported noise models include Gaussian, Poisson, quasipoisson, NB1/NB2 and both Gamma parameterizations, with a compound-count criterion, unpenalized spline backgrounds, temporal noise calibration and optional channel preclustering followed by full refitting. Conservative broad-profile fusion reunites complementary ion sets on nearly proportional profiles, with original-channel criterion checks and one final native likelihood refit under hard log-concavity constraints.

  • GCMSdeco reads ANDI-MS chromatograms, estimates time-varying PSFs with optional alkane calibration, fits a spline background and log-concave compounds, and exports all-channel spectra as NIST-compatible MSP files. Excel exports include sample identity, integrated areas and unit-L1 spectra. Compact fit objects reload observed CDF intervals for heatmaps and multi-panel chromatogram plots; recovered spectra and PDF reports are also supported.

  • Blind deconvolution supports smoothly time-varying PSFs with nwindows, and returns the corresponding sparse design as X for downstream deconvolution. Standard calibration uses full local peak windows and automatic PSF support; blind PSF updates retain positive peak tails alongside the fitted background. Blind fitting compares full-tail and concentrated PSF proposals by refitting the original observations and evaluating the requested information criterion.

  • Fast sparse generalized linear models with adaptive ridge and IRLS, including Gaussian, Poisson, binomial, Gamma and inverse Gaussian families and their links.

  • Negative binomial regression with estimated shape, joint multinomial regression in centered log-ratio coordinates, and Cox proportional hazards models.

  • Exact fixed-parameter NB1, common-scale Gamma and censored Gamma models, with native likelihood scores, tuning, screening and likelihood-ratio clustering. Quasipoisson models expose scaled log-likelihood kernels and quasi-information criteria through the regular scoring methods.

  • Conditional boundary polishing for relaxed nonnegative models removes likelihood-supported zero coefficients without a coefficient-size cutoff. Native Gaussian/Poisson kernels retain dense/sparse storage and Poisson OpenMP; other GLMs, negative binomial and Cox use their family-specific objectives.

  • Single and multichannel fitting, grouped predictor-row selection, and relaxed refitting for accurate coefficient estimates on the selected support.

  • Dense and sparse matrices, tall and wide designs, fast Eigen solvers and a native OpenMP backend for nonnegative identity-link Poisson models.

  • Sure Independence Screening (SIS) and Iterative Sure Independence Screening (ISIS), with compiled OpenMP screening and automatic ISIS when p_eligible > 10*n, excluding mandatory terms and predictors excluded by zero starts in every response channel. The same rule uses the training row count in each L0glm.tune() fit. Screening uses physical cores by default, with explicit thread-count overrides.

  • Small constrained solves for nonnegative compulsory terms and smooth spline backgrounds, alternating with sparse selection on the original channels. Gaussian, Poisson, other scalar GLMs and negative-binomial/Cox fits retain their distribution-specific likelihoods, dispersion and validation controls.

  • Fixed-PSF and blind deconvolution use positive signal-based starts with backgrounds, channel-specific PSF offsets, and synchronized final estimates. Automatic identity-link initialization accounts for nonnegative compulsory backgrounds and known offsets while preserving original coefficient units.

  • Observation weights, offsets, intercepts, mandatory predictors, normalization, starting values, nonnegativity constraints and validation-based early stopping.

  • Information-criterion presets and lambda-path tuning, five-fold cross-validation, one-standard-error models and coefficient-path plots. preset.lambda() returns the numeric AIC, BIC, GIC, EBIC or mBIC penalty directly.

  • sparseglm() comparisons with supported sparse-regression backends, common likelihood and support-recovery metrics, prediction and summary methods, mirrored deconvolution plots and multichannel heatmaps.

  • Reproducible simulations, large-scale examples and a documentation website with tutorials, measured benchmarks and a mathematical algorithm appendix.

  • Size-adjusted Poisson channel distances use a native OpenMP kernel with packed distance storage and scalar, row or channel-specific weights.

  • Native scalar-family deviance and Bregman clustering use explicit pooled likelihood merge costs, IC-scored tree cuts and dispersion-aware precision. Poisson aggregation preserves shared case weights.

  • Optional Poisson channel preclustering screens aggregated responses and refines on the original channels; aggregate-only fits retain the requested penalty. Starts, weights, offsets and original predictor dimensions are preserved.

  • Preclustering-aware lambda paths reuse training-derived memberships while estimating every model independently; cross-validation learns memberships and size factors within each training fold.