New distribution() plots the exposure-weighted distribution of the target variable as a plotly bar chart. It mirrors one_way()’s interface (bare or quoted column names, split, bins, time_unit, type_agg, ret) and reuses its binning and NA handling; a modelblueprint method pulls the target and exposure from the blueprint’s slots.
modelblueprint 1.6.3
Bug fixes
Constructing or validating a modelblueprint no longer fails on R < 4.3 with trying to get slot ... that is not an S4 object. S7 only exports its @ operator on R < 4.3 (base R handles S7 @ natively from 4.3), and the package did not import it, so every property access inside package code fell through to base’s S4 slot operator on older R.
Infrastructure
Declared minimum R version raised from 4.1.0 to 4.2.0 to match what CI actually verifies (an R 4.2.2 job mirroring the Azure ML deployment target).
H2O tests are skipped on CI runners, where the H2O JVM is unstable, and run in a weekly scheduled workflow instead (set MB_RUN_H2O_TESTS=1 to force them on).
modelblueprint 1.6.2
New features
gain(), pred_vs_obs() and residuals_grouped() on a modelblueprint now default to all available sets (train/test/holdout) instead of silently using only train. With more than one set they return a named list with one result per set; a single set returns the bare plot as before.
Bug fixes
save_plots() with selfcontained = FALSE now writes dependency links that resolve correctly. Previously the dependency folder was created nested inside the output directory a second time, and plots rendered blank once the output tree was moved or cleaned.
save_plots() gains a shareable libdir: model_validation() uses it so all HTML files in each output subdirectory share a single lib/ dependency folder instead of one folder per file.
one_way() on a modelblueprint with multiple variables omits skipped variables (e.g. >2,000 unique non-numeric levels) from the returned list instead of including NULL entries that broke save_plots().
modelblueprint 1.6.1
Bug fixes
residuals_grouped() now computes bin midpoints directly from the numeric break points instead of re-parsing cut() labels. Previously, labels in scientific notation (small prediction rates) produced meaningless midpoints.
gain() drops rows with missing obs/pred/exposure values (with a warning) before building the cumulative curve. Previously a single NA truncated the curve silently from that row onward.
one_way() and pdp() weighted bin means are no longer deflated when the target contains NAs: each mean’s denominator now counts only the exposure of rows where that value is non-missing.
model_predict() rejects H2O multinomial models with a clear error instead of silently returning factor level codes.
pred_vs_obs() and residuals_grouped() default methods fall back to unit weights when the exposure column is absent (matching one_way() and gain()), and validate the obs/pred columns up front.
pdp() on a modelblueprint keeps the @offset_name column when narrowing the working dataset, so models fit with offset() predict correctly.
predict() on an mb_seq with return_all = TRUE no longer returns duplicated columns, and mb_layer() rejects blueprints with duplicate @yhat_names (the later one silently overwrote the earlier).
The x-axis ordering of interval labels now understands scientific notation.
Gains charts support more than 12 competing scores (palette interpolation).
The example xgboost blueprints use the current xgboost() interface; the suggested xgboost version floor is now >= 3.0.0.
Improvements
Zero-exposure rows are replaced with @expo_0_rep (with a warning) before the rate division in gain(), pred_vs_obs(), and residuals_grouped() on a modelblueprint — previously the property was documented but unused and zero exposure produced Inf rates.
pdp() and one_way() gained a verbose argument (default FALSE); per-call progress and NA-relocation messages are now opt-in.
Hover text no longer shows Inf%/NaN% when a reference mean is zero.
filter(), mutate(), and left_join() generics are re-exported, so the modelblueprint methods work without attaching dplyr.
savemb() on an mb_seq fails with an informative message (serialisation of sequences is not yet supported).
