HNPclassifier 0.2.0
Breaking changes
hnp_umbrella() has a new interface. It now takes
separate X and Y arguments together with
importance_order instead of a single data frame
S and a class_col. It supports an arbitrary
number of ordered classes (T >= 2) rather than only
ternary classification.hnp_summary() now takes classifier,
X, Y and importance_order instead
of data and class_col, and accepts classifiers
that return class labels, probability matrices or score matrices, as
well as fitted model objects.hnp_map_classes() now accepts a variable number of
class labels via ... (in decreasing priority order) instead
of the fixed class_1, class_2,
class_3 arguments.- Removed
probability_to_score_1(),
probability_to_score_2(), hnp_umbrella_flex()
and hnp_box_plot().
New features
- Generalized the H-NP umbrella algorithm to any number of ordered
classes (
T >= 2).
- Added support for user-supplied pretrained models and pre-computed
score matrices in
hnp_umbrella() via the
pretrained_model and input_is_score
arguments.
- Added grid search over candidate thresholds
(
grid_search, grid_set, max_grid,
max_combinations) to minimize the weighted
misclassification objective, with a recursive multi-class threshold
search.
- Added
hnp_boxplot() to visualize and summarize
under-classification and overall error from confusion matrices,
supporting single- and two-method comparisons.
- Added data-generation helpers for examples and simulation
(
gen_data(), gen_normal_data(),
generate_ball_data(), etc.) and a neural-network scoring
helper (train_nn_and_get_scores()).
Improvements and bug fixes
hnp_delta_search() now uses
stats::pbinom() instead of an explicit
choose()-based summation, avoiding numerical overflow and
underflow for large samples.hnp_upper_bound() is now vectorized over the score
functions and adds validation for non-finite, NA and
length-mismatched score outputs.- Added stricter input validation across the public functions (class
labels, control levels, tolerances and split ratios).
- Added internal label mapping so that original class labels are
preserved in predictions and summaries.