Skip to main content
Use this checklist when reporting FPDE experiments, publishing results, or comparing attribution settings across runs.

​
Package and environment

  • Record the FPDE package version.
  • Record the Python version. FPDE requires Python 3.12 or newer.
  • Record dependency versions, especially NumPy and scikit-learn.
  • State whether FPDE was installed from PyPI or from a local checkout.
  • Preserve the commit hash when using a local checkout.

​
Data and preprocessing

  • Identify the dataset and any dataset version or download date.
  • Record the train, validation, and test split procedure.
  • Record all random seeds used for splitting, model fitting, and sampling.
  • Describe feature preprocessing, including scaling, encoding, imputation, and feature selection.
  • Confirm that FPDE receives data in the same feature space used by the classifier.
  • Record feature names and feature order.

​
Model

  • Name the classifier and implementation library.
  • Record model hyperparameters.
  • Record the fitted model version or artifact location.
  • Confirm that the model exposes predict_proba and classes_ when using FPDEEngine, grid search, lambda selection, or perturbation curves.
  • Report model performance on the evaluation split used for explanations.

​
FPDE configuration

  • Record the FPDE mode: Diff-FPDE, Cos-FPDE, or Hyb-FPDE.
  • Record lambda_hyb for Hyb-FPDE.
  • Record whether lambda_hyb was fixed manually or selected with FPDEEngine.select_lambda.
  • Record the lambda_hyb_grid used for validation selection.
  • Record normalize, usually "l1" or "none".
  • Record anchor_strategy, usually "mean", "zero", or "none".
  • Record eps for cosine computations.
  • Record the baseline vector used for deletion and insertion curves.
  • Record the validation fractions used by select_lambda or perturbation_curves.

​
Prototype state

  • Record how prototypes were built. FPDE currently provides class-mean prototypes.
  • Record the training data used to build prototypes.
  • Record class labels and label ordering.
  • If you use prepare_fpde_context, preserve the context-generation inputs.

​
Explanation outputs

  • Save raw attribution vectors before visualization normalization.
  • Save target and rival labels for each explanation.
  • Save evidence values and exactness residuals.
  • Save positive_score and negative_score when comparing contrasts.
  • Save grid-search rows or validation-selection rows when using automated selection.

​
Validation and sanity checks

  • Run the test suite with python -m pytest.
  • Verify that attribution shapes match the number of features.
  • Check that attribution values are finite.
  • Check that target and rival labels are different for multiclass explanations.
  • Inspect deletion and insertion curves for a representative sample.
  • Confirm that repeated runs with fixed seeds are deterministic.

​
Minimal reporting template

FPDE version:
Python version:
Model:
Dataset:
Preprocessing:
Train/validation/test split:
Prototype construction:
FPDE mode:
lambda_hyb:
normalize:
anchor_strategy:
eps:
baseline:
Validation fractions:
Random seeds:
Repository commit: