Skip to contents

Changelog

psvr 0.0.2.9003 (development)

Internal changes

  • .smo_solve() (R/smo_solve.R) now implements Theorems 3 and 8 of arXiv:2605.01446 v3:

    • Theorem 3 (asymmetric freeze): the uniform freeze-counter threshold n_freeze = 5L is replaced by per-sample, per-variable-type thresholds:

      n_freeze_astar_per[k] = max(1L, floor(n_freeze * y[k] / mean(y)))
n_freeze_alpha_per[k] = max(5L, ceil( n_freeze * mean(y) / y[k]))

      α*-variables tied to high-y_k samples freeze faster; α-variables tied to high-y_k samples freeze slower. Reduces to the scalar default when y is homogeneous.

    • Theorem 8 (per-pair tolerance): the uniform convergence threshold tol * mean(y) is replaced at the convergence test by a per-pair threshold tol * max(y[p], y[k_j_w1]), where (p, k_j_w1) is the WSS1 pair. The WSS3 candidate filter at the working-set selection step retains the global tol * mean(y) for noise-floor purposes.

Behavior change

  • SMO iteration count decreases by ~10–30% on heterogeneous-target datasets (rho_y = max(y) / min(y) >= 50). The wall-clock speedup is smaller (depends on N: kernel-matrix construction overhead dominates at moderate N). Empirical benchmark at N = 200, rho_y = 1273: 24.7% iter reduction, 5.6% wall-clock speedup. Homogeneous datasets (rho_y ~ 1) are unaffected — the new thresholds collapse to the F3 defaults.

Numerical drift

  • Predictions on heterogeneous-target datasets differ from v0.0.2.9002 by O(tol * max(y)). Empirical magnitude on the snapshot fixture (rho_y = 6.7, max(y) = 2.6, tol = 1e-3): max drift 8.5e-4 per prediction (well below the per-pair tolerance floor of 2.6e-3).
  • The 28 golden snapshot tests have been re-recorded with F4 numerics as the new regression baseline at tolerance 1e-10 within F4. Bit-identicalidad with v0.0.2.9002 is intentionally NOT preserved on SMO-backed paths; LS-SVR paths (Models 3, 4 via linear system) remain bit-identical.

Paper deviation

  • The convergence test uses the WSS1 pair (i_w1, j_w1) rather than the paper’s literal text (“j* = WSS3 pick”). Rationale documented inline in R/smo_solve.R: Delta_w3 <= Delta_w1 by construction, so testing Delta_w3 against the tolerance would stop prematurely before the true KKT optimality gap is below tolerance. Flagged for paper-side notation fix in F8 (paper TODO #4).

psvr 0.0.2.9002 (development)

Internal changes

  • .fit_mape_sym() (Model 2) now invokes Algorithm 2 (Theorem 2 of arXiv:2605.01446 v3) via the new .adaptive_spectral_shift() in R/kernel-spectral.R. Two-pass shifted power iteration estimates λ_min(Ωs); if numerically negative, an adaptive μ·I shift is applied so the SMO Hessian is provably PSD. Diagnostics (mu, lambda_min_hat, lambda_max_hat, branch_taken, n_power_iterations) surface under psvr_fit$solver_meta$spectral for symmetric MAPE fits; NULL for Models 1, 3, 4. The implementation deviates from the paper’s Algorithm 2 line 6 by using the spectral radius |Pass 1 Rayleigh| rather than the signed Rayleigh as the Pass 2 shift; rationale is documented inline in R/kernel-spectral.R and to be flagged as a paper-side erratum in F8. With the three Mercer kernels supported by make_kernel() (RBF / linear / polynomial), Ωs is always PSD by Aronszajn closure / Schur product theorem, so the shifted branch is unreachable in production fits and predictions are bit-identical to v0.0.2.9001 at tolerance 1e-10 on the 16 + 12 golden snapshots.

psvr 0.0.2.9001 (development)

Internal changes

  • The SMO solver now reads kernel values via a .make_kernel_accessor() wrapper (R/kernel-accessor.R), decoupling the inner loop from the matrix representation. The F2 implementation is a thin closure over the materialised Ω/Ωs; future phases (F3 spectral shift, F6 LRU cache, F7 block working set) will swap in alternative implementations without touching the SMO loop or the fitters. No user-visible change; predictions are bit-identical to v0.0.2.9000 at tolerance 1e-10 on the 16 + 12 golden snapshots.

psvr 0.0.2.9000

New features

  • New unified entry point psvr(X, y, loss, sym, kernel, ...). Selects among the four model families via loss = "mape" | "rmspe" and sym = NULL | +1L | -1L. Returns a single psvr_fit S3 object with a unified field schema (alpha, b, support_data, support_targets, n_train, n_sv, p_train, hyperparameters, solver_meta, plus loss, sym, kernel).
  • New methods on the psvr_fit class: predict(), print(), coef(), and summary(). The summary method is new — there was no analogue on the four legacy classes.

Deprecations

  • mape_svr(), mape_sym_svr(), rmspe_lssvr(), rmspe_sym_lssvr() are soft-deprecated. They continue to work and return the same legacy shape (psvr_mape, psvr_mape_sym, psvr_rmspe, psvr_rmspe_sym) with the same predict/print/coef methods, but each emits .Deprecated("psvr"). Migrate by replacing, for example, mape_svr(X, y, kernel = K, C = 10, eps = 5) with psvr(X, y, loss = "mape", kernel = K, C = 10, eps = 5). Scheduled removal: v0.2.0 or later.

Internal changes

  • The 12 parsnip fit wrappers (psvr_mape_rbf_fit() etc.) are now tagged @keywords internal and hidden from the pkgdown reference index. They remain exported (parsnip’s resolver requires it) but are no longer advertised as user API; call [psvr()] for direct fitting. The 12 spec constructors (psvr_mape_rbf() etc.) remain public API and are unchanged.
  • The four model fitter bodies have been renamed to internal helpers (.fit_mape(), .fit_mape_sym(), .fit_rmspe(), .fit_rmspe_sym()). The deprecation wrappers delegate to these directly.
  • DRY consolidation: shared validation, large-N warnings, predict dispatch, preconditioner resolution, and kernel description live in new internal-only files under R/utils-*.R.
  • Model 2 (symmetric MAPE epsilon-SVR) now uses sym_kernel_matrix() to build Ωs = ½(Ω + a·Ω*) (matching Model 4’s existing convention), with diagonal jitter 0.5e-6 to preserve bit-identicality with the pre-F1 path.

psvr 0.0.2 (2026-04-30)

New features

  • The precondition argument from rmspe_lssvr() and rmspe_sym_lssvr() is now configurable via parsnip’s set_engine() for all six RMSPE spec functions (rbf/poly/linear, both symmetric and non-symmetric). Default is "auto"; pass via set_engine("psvr", precondition = "always"), "never", or a numeric threshold. Not registered as a tunable parameter — it is a configuration flag, not a hyperparameter.

psvr 0.0.1

New features

  • Added optional preconditioner for LS-SVR variants (rmspe_lssvr(), rmspe_sym_lssvr()) via the precondition argument. Accepted values: “always”, “never”, “auto” (default; activates when max(y)/min(y) > 10), or a numeric threshold. The preconditioner is a mathematically exact change of variable that improves numerical conditioning at large target dynamic ranges without changing predictions in exact arithmetic.
  • Returned model objects now include precondition_applied (logical) for diagnostic transparency. The print method displays this field when TRUE.

Documentation

  • Updated @details for both LS-SVR fitters with the bordered-system formulation and the recovery formula α_k = α̃_k / y_k.

psvr 0.0.0.9005

New features

  • cost_psvr_ls_data(y, n, width_log2) — a data-driven cost range for LS-SVR psvr models (m3, m4). Default upper bound scales as log2(var(y) * N) + width_log2 (default width_log2 = 4), the standard heuristic for the LS-SVR regularisation parameter Γ (Suykens et al. 2002, Least Squares Support Vector Machines, §3.1.3). Use this instead of cost_psvr() for m3/m4 workflows; the static [-2, 10] log2 range underestimates Γ on benchmark datasets (Boston Housing optimum Γ ≈ 1.7 × 10⁴, vs. cost_psvr() upper bound 2^10 = 1024).
  • psvr_option_add_cost_ls(wf_set, y, ...) — a convenience wrapper that applies cost_psvr_ls_data() to every LS-SVR psvr workflow in a workflow set (those whose wflow_id matches m3|m4), analogous to psvr_option_add() for rbf_sigma.

psvr 0.0.0.9004

New features

  • sym_type is now a tunable parsnip argument in psvr_mape_sym_rbf() and psvr_rmspe_sym_rbf(). Pass sym_type = tune() to let CV select between even (a = 1) and odd (a = -1) symmetry automatically.
  • Engine default a = 1L removed from symmetric RBF specs; symmetry type is now controlled exclusively through the sym_type model argument.

psvr 0.0.0.9003

UX improvements

  • New function rbf_sigma_psvr_data(): combines sigma_heuristic() and rbf_sigma_psvr() in a single call, returning a dials parameter with a data-driven search range centred on the median pairwise distance.

  • New function psvr_option_add(): applies option_add() to every psvr workflow (those whose wflow_id matches m1|m2|m3|m4) in a workflow set simultaneously, replacing the rbf_sigma parameter with a data-driven one. Replaces four individual option_add() calls in typical workflows.

  • Symmetric model specs (psvr_mape_sym_*, psvr_rmspe_sym_*): even symmetry (a = 1L) is now the default engine argument. Calling set_engine("psvr") without a = 1L now produces the same result as before — no action required for existing code that already passed a = 1L.

  • All four core fitting functions now emit a warning when N > 2000 to remind users that the O(n²) kernel matrix may be expensive.

  • The y > 0 input check in all four core fitting functions now reports the number of offending values and their minimum, making it easier to diagnose data issues.

psvr 0.0.0.9002

User-friendly hyperparameter defaults

  • New function margin_percentage(): a dials parameter for the epsilon tube half-width in MAPE models, expressed in percentage units with default range [1, 20]. Replaces the absolute-unit dials::svm_margin() for all 6 MAPE specs.

  • New function sigma_heuristic(): returns the median pairwise Euclidean distance of a predictor matrix — a standard data-driven starting point for the RBF kernel bandwidth (Schölkopf & Smola, 2002).

  • New function rbf_sigma_psvr(): a dials parameter for the RBF kernel bandwidth with a finalize hook that auto-sets the search range using sigma_heuristic() when training data are available. Replaces dials::rbf_sigma() for all RBF specs.

  • New function cost_psvr(): a dials parameter for the regularisation parameter with range [-2, 10] on the log2 scale (approx. 0.25 to 1024), wider than dials::cost() to accommodate the larger values typically needed by LS-SVR models. Replaces dials::cost() for all 12 specs.

  • rbf_sigma_psvr(): removed non-functional finalize hook. The search range must be set manually using sigma_heuristic() and option_add(). See ?rbf_sigma_psvr for the recommended workflow.

psvr 0.0.0.9001

tidymodels / parsnip integration — breaking change

Expanded parsnip integration from 4 to 12 model specs following the tidymodels pattern used by svm_rbf(), svm_poly(), and svm_linear():

Kernel parameters are now tunable parsnip args mapped to existing dials params: rbf_sigma → dials::rbf_sigma(), degree → dials::degree(), scale_factor → dials::scale_factor(). The kernel closure is built inside each fit wrapper and no longer appears in the parsnip layer.

The old single-spec API (psvr_mape(), psvr_mape_sym(), psvr_rmspe(), psvr_rmspe_sym()) has been removed. Migrate by replacing, for example, psvr_rmspe(cost = tune()) |> set_engine("psvr", kernel = K) with psvr_rmspe_rbf(cost = tune(), rbf_sigma = 1) |> set_engine("psvr").

Documentation and testing

  • All vignettes and pkgdown articles updated to the new spec-based API.
  • 12 new smoke tests (fit + predict) for all specs in tests/testthat/test-parsnip.R.
  • pkgdown articles reorganised into three named groups: Get Started, Case Studies, and Technical Notes.

psvr 0.0.0.9000

Initial development release.

New models

  • mape_svr() — epsilon-SVR with MAPE loss (Model 1). Solves the dual QP via osqp with per-sample box constraints |βk| ≤ 100C/yₖ.

  • mape_sym_svr() — symmetric epsilon-SVR with MAPE loss (Model 2). Enforces f(x) = a·f(-x) by replacing the kernel with the symmetric kernel Ks(xi, xj) = K(xi, xj) + a·K(xi, -xj).

  • rmspe_lssvr() — LS-SVR with RMSPE loss (Model 3). Solves the (N+1)×(N+1) linear system directly via base::solve().

  • rmspe_sym_lssvr() — symmetric LS-SVR with RMSPE loss (Model 4). Same linear system as Model 3 with the symmetrized kernel matrix Ωs = ½(Ω + a·Ω*).

All four models are derived from a unified percentage-error loss framework; see Benavides-Herrera et al. (2026) for the mathematical proofs.

Kernel interface

  • make_kernel() — factory returning a kernel closure from type ∈ {"rbf", "linear", "polynomial"}. RBF and even-degree polynomial kernels satisfy the symmetry assumption (Assumption 3) required by Models 2 and 4.

tidymodels / parsnip integration

Four parsnip model specifications for use within tidymodels workflows:

  • psvr_mape() — spec for Model 1; hyperparameters cost and svm_margin.
  • psvr_mape_sym() — spec for Model 2; hyperparameters cost and svm_margin.
  • psvr_rmspe() — spec for Model 3; hyperparameter cost (maps to Γ).
  • psvr_rmspe_sym() — spec for Model 4; hyperparameter cost (maps to Γ).

The kernel and (for symmetric models) symmetry parameter a are engine arguments supplied via set_engine("psvr", kernel = ..., a = ...). Hyperparameters map to dials::cost() and dials::svm_margin() for compatibility with tune_grid().

Testing

  • 84 unit tests covering all four models: kernel correctness, QP/linear-system solutions, support vector selection, bias recovery, predict consistency, and input validation.