orbital 0.7.0

Breaking changes

• estimate_orbital_size() now errors for a workflow whose model it has no estimate for, rather than counting that model as zero characters and returning the recipe's size as the whole workflow's. The model is usually the bulk of the expression, so the number it returned could be off by orders of magnitude while looking ordinary, and it did not move as the model's hyperparameters changed. Use a model that has an estimate, or generate the expression with orbital() and measure it directly. (#167)

• orbital() now errors for a bare model fit, which was never documented input, rather than returning something that looked like a result. A regression fit handed in directly took the classification path whatever the model was, so orbital(rpart::rpart(mpg ~ ., mtcars)) came back as a character vector named orbital_tmp_class_name with no error and no warning. Fit the model with parsnip::fit(), or use a workflow, as the documentation has always described. (#113)

• orbital() now errors for type = "prob" on models that have no probability to give, rather than fabricating one. A decision value is uncalibrated, so putting it through a logistic would invent a calibration the model does not have. Use type = "class" for these models, or fit an engine that estimates probabilities. (#159)

New models

Trees and ensembles

• bag_tree() with the "rpart" and "C5.0" engines is now supported for type = "class". These models vote over their ensemble and expose no probability, so type = "prob" is refused. (#161)

• bart() with the "dbarts" engine is now supported for regression. Classification is refused, since it uses a probit link that cannot be translated. (#162)

• boost_tree() with the "C5.0" engine is now supported for type = "class", including multi-trial boosting. (#161)

• boost_tree() with the "h2o_gbm" engine, and rule_fit() with the "h2o" engine, are now supported for regression and classification. A running H2O cluster is needed to build the orbital object, but not to use one afterwards. (#166)

• C5_rules() with the "C5.0" engine is now supported for type = "class". (#161)

• decision_tree() with the "C5.0" engine is now supported for type = "class". Its leaves carry a class label rather than class counts, so type = "prob" is refused. (#173)

• rand_forest() with the "aorsf" engine is now supported for regression. Classification is refused, since aorsf votes across the forest and exposes no probability. Note that aorsf splits on observed linear-combination values, so a row that lands exactly on a split boundary can take the other branch than predict() did. (#173)

• rand_forest() with the "partykit" engine is now supported for regression. (#161)

• rule_fit() with the "xrf" engine is now supported for regression and for binary classification. Multiclass outcomes are refused, since xrf only fits Gaussian and binomial models. (#164)

Discriminant analysis

• discrim_linear() and discrim_quad() with the "MASS" engine are now supported for type = "class" and type = "prob". (#160)

• discrim_linear() with the "mda", "sda", and "sparsediscrim" engines is now supported for type = "class" and type = "prob". (#163)

• naive_Bayes() with the "klaR" and "naivebayes" engines is now supported for type = "class" and type = "prob". Both engines default to usekernel = TRUE, which fits a kernel density per predictor and has no closed form; refit with usekernel = FALSE to translate one. (#163)

Linear models

• linear_reg() with the "glm" engine is now supported. (#160)

• logistic_reg() with the "LiblineaR" engine is now supported for type = "class" and type = "prob". (#164)

• multinom_reg() with the "nnet" engine is now supported for type = "class" and type = "prob". (#160)

Support vector machines

• svm_linear() with the "kernlab" engine is now supported for regression and classification. (#164)

• svm_linear() with the "LiblineaR" engine is now supported for regression, in addition to the type = "class" support added in #159. (#164)

Other models

• mlp() with the "nnet" engine is now supported for regression and classification. (#160)

• null_model() is now supported for regression and classification. (#160)

• pls() with the "mixOmics" engine is now supported for regression and for type = "prob". type = "class" is refused, since mixOmics assigns a class by distance to the class centroid rather than by the largest per-level value. (#163)

Improvements

• orbital(separate_trees = TRUE) now works for rand_forest() with the "aorsf" and "partykit" engines. The argument used to be accepted and silently ignored for every model orbital has no method of its own for; it is now honored for any regression ensemble whose per-tree expressions tidypredict exposes. (#173)

• orbital() now supports classification models that reach the tidypredict fallback, rather than refusing them. This covers multiclass probability models such as MASS::lda(), models returning an uncalibrated decision value such as LiblineaR SVMs, and models predicting a class label directly such as C50::C5.0(). (#159)

Bug fixes

• mars() models with the "earth" engine now generate 1 / (1 + exp(-x)) for binary classification, rather than the equivalent 1 - 1 / (1 + exp(x)). Predictions are unchanged. (#158)

• orbital() no longer falls back to tidypredict when one of its own model methods errors. An error in a native method was previously caught and silently replaced with a tidypredict result, so a bug could still produce an answer. Whether a native method exists is now checked directly, and errors from it propagate. (#156)

• orbital() now uses the model's own class order for binary probabilities, rather than assuming it matches the order of the outcome's factor levels. Every engine but h2o orders them the same way, so only h2o models were affected, and only when the outcome's levels were not in sorted order; for those both probability columns were swapped and the class inverted. (#166)

• orbital() now returns the correct classes for svm_linear() models with the "kernlab" engine. kernlab classifies by the sign of its decision function and calibrates its probabilities separately, so cutting those probabilities at 0.5 disagreed with the model for rows near the boundary. (#164)

• orbital() now returns the correct classes for svm_linear() models with the "LiblineaR" engine. The sign of the decision value was read as meaning the second outcome level, but LiblineaR orients it by its own class order, which need not match the order of the outcome's factor levels. When the two disagreed every class was inverted. (#164)

• orbital(separate_trees = TRUE) now returns NA for rows with a missing predictor, matching what separate_trees = FALSE has always returned. The individual tree expressions fall through to their default branch when a split variable is NA, so such rows previously received a confident-looking prediction computed from no usable data. (#158)

• orbital(separate_trees = TRUE) now applies CatBoost's scale and bias to binary classification models. The regression path applied them and the binary path did not, so the two disagreed with separate_trees = FALSE whenever a model carried a non-default scale or bias. (#158)

• print() no longer corrupts numbers when rounding them for display. Numbers such as 6.75044994983228 and -0.0901719835820594 were printed as 6.750.5 and -017198. Only the printed output was affected; the expressions themselves were always correct. (#155)