NEWS.md

In hstats: Interaction Statistics

hstats 1.1.2

ICE plots

• The ICE plot of a multioutput model without BY variable will now be using facets (instead of color). Use swap_dim = TRUE for the old behavior.

API

• {mlr3}: Non-probabilistic classification now works.
• {mlr3}: For probabilistic classification, you now have to pass predict_type = "prob".

hstats 1.1.1

Performance improvements

• For pure data.frames (no tibbles, data.tables etc.), most functions are significantly faster (#110).
• Slight speed-up of permutation importance for non-matrix X (#109).

Other changes

• In multivariate cases, it was possible that normalized H-statistics could equal 0/0 (= NaN). Such values are now replaced by 0 (#107).
• Removed an unnecessary special case when calculating column means (#106).

hstats 1.1.0

Enhancements

• {hstats} now also works for factor predictions. The levels are represented by one-hot-encoded columns (PR#101).
• The plot method of a two-dimensional PDP has recieved the option d2_geom = "line". Instead of a heatmap of the two features, one of the features is moved to color grouping. Combined with swap_dim = TRUE, you can swap the role of the two v variables without recalculating anything. The idea was proposed by Roel Verbelen in issue #91, see also issue #94.

Bug fixes

• Using BY and w via column names would fail for tibbles. This problem was described in #92 by Roel Verbelen. Thx!

Other changes

• Much faster one-hot-encoding, thanks to Mathias Ambühl (PR#101).
• Most functions are slightly faster (PR#101).
• Add unit tests to compare against {iml}.
• Made all examples "tibble" and "data.table" friendly.
• Revised input checks in loss functions (relevant for perm_importance() and average_loss()).

hstats 1.0.0

Major changes

• Quantile approximation: hstats() now has the option approx = FALSE. Set to TRUE to replace values of dense numeric columns by grid_size = 50 quantile midpoints. This will bring a massive speed-up for one-way calculations. Use this option when one-way calculations are slow, or when you want to increase n_max.
• hstats(): n_max has been increased from 300 to 500 rows. This will make estimates of H-statistics more stable at the price of longer run time. Reduce to 300 for the old behaviour.
• hstats(): Three-way interactions are not anymore calculated by default. Set threeway_m to 5 for the old behaviour.
• Revised plots: The colors and color palettes have changed and can now also be controlled via global options. For instance, to change the fill color of all bars, set options(hstats.fill = new value). Value labels are more clear, and there are more options. Varying color/fill scales now use viridis (inferno). This can be modified on the fly or via options(hstats.viridis_args = list(...)).
• "hstats_matrix" object: All statistics functions, e.g., h2_pairwise() or perm_importance(), now return a "hstats_matrix". The values are stored in $M and can be plotted via plot(). Other methods include: dimnames(), rownames(), colnames(), dim(), nrow(), ncol(), head(), tail(), and subsetting like a normal matrix. This allows, e.g, to select and plot only one column of the results.
• perm_importance(): The perms argument has been changed to m_rep.
• print() and summary() methods have been revised.
• The arguments w (case weights) and y (response) can now also be passed as column names.

Minor changes

• Statistics: The argument top_m has been moved to the plot() method.
• Statistics: The clipping threshold eps of squared numerator statistics has been reduced from 1e-8 to 1e-10. It is now handled in hstats() instead of the statistic functions.
• H-squared: The $H^2$ statistic stored in a "hstats" object is now a matrix with one row (it was a vector).
• pd_importance(): The "hstats" object now contains pre-calculated PD-based importance values in $pd_importance.
• summary.hstats() now returns an object of class "hstats_summary" instead of "summary_hstats".
• average_loss() is more flexible regarding the group BY argument. It can also be a variable name. Non-discrete BY variables are now automatically binned. Like partial_dep(), binning is controlled by the by_size = 4 argument.
• average_loss() also returns a "hstats_matrix" object with print() and plot() method. The values can be extracted via $M.
• The default v of hstats() and perm_importance() is now NULL. Internally, it is set to colnames(X) (minus the column names of w and y if passed as name).
• Missing grid values: partial_dep() and ice() have received a na.rm argument that controls if missing values are dropped during grid creation. The default TRUE is compatible with earlier releases.
• Missing values in hstats(): Discrete variables with missings would cause rowsum() to launch repeated warnings. This case is now catched.
• The position of some function arguments have changed.
• perm_importance(): The default of verbose is TRUE again.

hstats 0.3.0

This is intended to be the last version before 1.0.0.

Visible changes

• Grid of ice() and partial_dep(): So far, the default grid strategy "uniform" used pretty() to generate the evaluation points. To provide more predictable grid sizes, and to be more in line with other implementations of partial dependence and ICE, we now use seq() to create the uniform grid.
• h2_pairwise() and h2_threeway() will now also include 0 values. Use zero = FALSE to drop them, see below. The padding with 0 is done at no computational cost, and will affect only up to pairwise_m and threeway_m features.
• The print() method of summary.hstats() is less verbose.

Improvements

• h2_overall(), h2_pairwise(), h2_threeway(), plot.hstats(), and summary.hstats() have received an argument zero = TRUE. Set to FALSE to drop statistics having value 0.
• perm_importance() and average_loss() will now recycle a univariate response when combined with multivariate predictions. This is useful, e.g., when the prediction function represents the predictions of multiple models that should be evaluated against a common response.

Bug fixes

• All progress bars were initialized 1 step too late.
• perm_importance() and average_loss() would fail for "mlogloss" in case the response y was univariate and non-factor/non-character.

Other changes

• All available H-statistics are now calculated within hstats() and attached to the resulting object. Each statistic is stored as list with numerator and denominator matrices/vectors. The functions h2(), h2_overall(), h2_pairwise(), and h2_threeway(), print.hstats(), summary().hstats(), plot.hstats() will use these without having to recalculate the required numerators and denominators. The results, however, are unchanged.

hstats 0.2.0

New major features

• average_loss(): This new function calculates the average loss of a model for a given dataset, optionally grouped by a discrete vector. It supports the most important loss functions (squared error, Poisson deviance, Gamma deviance, Log loss, multivariate Log loss, absolute error, classification error), and allows for case weights. Custom losses can be passed as vector/matrix valued functions of signature f(obs, pred). Note that such a custom function needs to return per-row losses, not their average.

• perm_importance(): H-statistics are often calculated for important features only. To support this workflow, we have added permutation importance regarding the most important loss functions. Multivariate losses can be studied individually or collapsed over dimensions. The importance of feature groups can be studied as well. Note that the API of perm_importance() is different from the experimental pd_importance(), which is calculated from a "hstats" object.

Major changes in defaults

• hstats() now uses the default feature vector v = colnames(X), simplifying the API in most cases. The typical call is now hstats(object, X = Feature data).
• h2_overall(), h2_pairwise(), h2_threeway(), pd_importance() by default do not plot results anymore. Set plot = TRUE to do so.

Minor changes

• summary.hstats() now returns an object of class "summary_hstats" with its own print() method. Like this, one can use su <- summary() without printing to the console.
• The output of summary.hstats() is printed slightly more compact.
• plot.hstats() has recieved a rotate_x = FALSE argument for rotating x labels by 45 degrees.
• plot.hstats() and summary.hstats() have received explicit arguments normalize, squared, sort, eps instead of passing them via ....
• plot.hstats() now passes ... to geom_bar().
• Slight speed-up of hstats() in the one-dimensional case.

Bug fixes

• Probabilistic {mlr3} classifiers did not work out-of-the box. This has been fixed.

hstats 0.1.0

This is the initial release.

hstats documentation built on May 29, 2024, 6:43 a.m.