Changes in version 0.4.0 (2026-05-11)                  

New features

  - Added interactions(), which ranks numeric predictor pairs by the
    count-weighted magnitude of their centred second-order ALE surfaces.
  - Added top_n to bivariate() for method = "ale" so the plotting
    interface can rank all eligible pairs with interactions() and
    display only the strongest interaction surfaces.

                        Changes in version 0.3.1                        

Fixes

  - bivariate(method = "ale") now masks grid cells that contain no
    observations as NA in the returned surface and renders them in the
    fill scale's na.value (default light grey) rather than colouring
    them with a value extrapolated from neighbouring cells. This
    prevents the interaction surface from displaying confident estimates
    over regions of feature space that the data does not support, which
    previously misled readers when predictors were correlated.
  - bivariate(method = "ale") now defaults to n = 10 (the second-order
    surface uses an n x n cell grid; n = 40 left most cells empty for
    typical data) and to rug = TRUE so the data density is always
    visible.
  - ale_surface_limits() is robust to surfaces that are entirely NA.
  - Updated docs/Rd to describe the half-cell centring convention used
    for the 2D ALE accumulation and the empty-cell masking behaviour.
  - Enabled categorical ALE in multimodel() and mapcurve(), matching the
    univariate ALE support and using a shared level order across
    ensemble members when multimodel() averages unordered factor
    effects.
  - Added extrapolate = FALSE to bivariate() so unsupported ALE grid
    cells remain masked by default but can be shown on request using the
    interpolated values already used internally for accumulation.

                        Changes in version 0.3.0                        

New features

  - Extended univariate() with method = "profile", "pdp", "ice", and
    "ice+pdp" so single-profile, partial dependence, and ICE plots share
    one entry point.
  - Added method = "ale" to univariate() for accumulated local effects
    curves on numeric predictors.
  - Changed univariate(method = "ale") to warn and skip factor
    predictors instead of failing when numeric predictors are also
    available.
  - Split univariate() and mapcurve() sampling controls so n sets
    numeric grid resolution while background_n sets the number of
    randomly sampled background rows used for PDP/ICE.
  - Added interval to univariate() and mapcurve() so method = "pdp" can
    draw central quantile ribbons for numeric predictors.
  - Added adaptive raster sampling for univariate() so PDP/ICE methods
    can draw more background predictor combinations when predict_data
    comes from a SpatRaster.
  - Added bivariate() for bivariate response surfaces with static
    heatmap and contour views.
  - Extended bivariate() with method = "pdp" and method = "ale", plus
    optional marginal rugs for numeric predictor pairs in static plots.
  - Added optional interactive 3D response surfaces for numeric
    predictor pairs when plotly is installed.
  - Added support for selecting predictor pairs by column name or column
    index.
  - Added support for list-valued fun in multimodel() so mixed model
    types can use model-specific prediction wrappers before averaging
    curves.

Changes

  - Renamed the bivariate plotting mode argument from type to plot_type
    so model-specific type arguments can still be passed through ... to
    predict().
  - Changed bivariate(plot_type = "heatmap") to use a viridis fill scale
    by default and to stop drawing contour overlays on heatmaps.
  - Updated the random forest species distribution vignette to use a
    smaller predictor set and to include univariate() examples for
    profile, PDP, and ICE + PDP plots.
  - Stopped drawing connecting lines for unordered factor predictors in
    univariate(), so categorical panels no longer imply numeric
    intervals.