twinsvm 0.0.4
- Test suite no longer compares support-vector counts at a fixed
tolerance. The count thresholds dual variables that can sit exactly at
the cost bound, so it varies with the BLAS the checking machine links
against; this made
test-baselines.R fail on the CRAN MKL
and OpenBLAS check flavors while passing everywhere else. Counts are now
compared up to a small fraction of the sample size, and the exact-value
regression test is skip_on_cran(). Per-point prediction
agreement against e1071 is unchanged and remains the strict
correctness check.
- Reference-fit tolerances relaxed from 1e-6 to 1e-5 for the same
reason.
- No model-fitting or prediction behavior changed.
twinsvm 0.0.3
- New
resample_accuracy() evaluates tsvm()
and svms() over repeated stratified random holdout splits
and reports the mean test accuracy together with its between-split
standard deviation and standard error, so method comparisons can be
judged against sampling noise. Splits are drawn once and shared across
models for a paired comparison. Adds print and
plot (boxplot) methods for the returned
resample_accuracy object.
- The animation test for
morph_boundary() now skips
cleanly when the suggested gganimate package is absent, so
R CMD check passes without Suggests installed. No
model-fitting or prediction behavior changed.
twinsvm 0.0.2
- Fixed cross-platform test portability on macOS by avoiding bitwise
equality checks for generated floating-point data.
- No model-fitting or prediction behavior changed.