JuliaTrustworthyAI/CounterfactualExplanations.jl: v1.5.0

CounterfactualExplanations v1.5.0

Diff since v1.4.5

Version [1.5.0] - 2026-02-14

Breaking

• Public API (should be) unaffected, but some major changes to internal API to comply with new autodiff routines.

Removed

• Temporarily removed tests for NeuroTreeModels.jl pending updates to compat.
• Removed stale dependencies from docs.

Added

• Added a new evaluation measure feature_sensitivity(ce::AbstractCounterfactualExplanation) that can be used to compute the proposed absolute changes for features.
• Added an optional callback::Union{Nothing,Function} argument to generate_counterfactual to allow users to either display or store anything relevant from the full CounterfactualExplanation object. This is useful if the user has specified return_flattened=true but still needs something specific from CounterfactualExplanation that is not returned with FlattenedCE.

Changed

• Moved from Flux.jl for autodiff to DifferentiationInterface.jl. Not all backends working for basic examples, but considerable speedup for ForwardDiff compared to Zygote. The following backends work for generic generator:
  • Zygote.jl
  • ForwardDiff.jl
  • PolyesterForwardDiff.jl
  • GTPSA.jl
  • FiniteDiff.jl
  • Symbolics.jl
• Avoid aggregating divergence metrics across runs.
• Changed the way mutability constraints can be supplied: users can now supply a tuple of pairs of feature indices and there corresponding constraints.
• Removed a bug in reconstruct_cat_encoding that turned the counterfactual_state object from a matrix into a vector.
• Added option to provide additional keyword arguments to validity_strict(ce::CounterfactualExplanation; kwrgs...).
• Small bug fix for validity function.
• Improved find_potential_neighbours method.
• Added possibility to run compute_divergence for a fixed number of samples.

Merged pull requests:

• Ct and changes to differentiation (#518) (@pat-alt)
• docs (#519) (@pat-alt)

Closed issues:

• New Flux API causing issues (#517)