Skip to main content
Back to timeline
arXivSource publication:

FADO swaps correlation for intervention in tabular explanations, returning every feature's intervention effect in one forward pass

Synopsis

Explanation for tabular machine learning has largely meant predictive attribution, which asks whether a feature helps predict the outcome and captures correlation rather than causation. FADO instead trains a prior-fitted network on synthetic tables drawn from structural causal models, where the true intervention effects are known and serve as supervision, so at test time a single observational table yields the full vector of per-feature intervention effects in one forward pass. In settings where standard predictive attribution fails, including proxy variables, confounders, colliders and placebos, FADO produces a feature-importance ranking that differs from predictive attribution and drives non-causal but predictive features toward zero.

Source-provided article image: FADO: Learning Causal Explanations with Prior-Fitted Networks
arXiv

Interpretation

Explanation tools for tabular data can now answer whether intervening on a feature would change the outcome, not only whether the feature helps predict it. FADO defines each feature's intervention effect as a standardized contrast and uses it to rank causal feature importance, alongside or instead of predictive attribution such as SHAP. Earlier explanation work is mostly predictive attribution, such as SHAP and permutation importance, which measures a feature's contribution to a prediction and reflects association; answering an intervention question usually requires strong additional causal assumptions or a known causal graph. On five hand-built structural causal models, FADO assigns near-zero effects to predictive but non-causal features such as target descendants and colliders, while predictive attribution baselines report non-zero effects; on held-out random structural causal models it outperforms every other method in every comparison column except where LiNGAM's identifiability conditions hold.

The mechanism turns learning to explain into one amortized inference step: a transformer reads a table and outputs the whole causal feature profile. Prior-fitted networks had been used for causal effect estimation, but those methods require designating a single binary treatment column and are called once per feature; FADO designates no treatment and returns all feature effects in a single forward pass. Training data come from eleven structural-causal-model prior families, with closed-form labels for the linear families and Monte Carlo estimates from post-intervention models for the rest; the model is trained on tables of up to 15 features, with an appendix extension to 64 features and an interactive browser demo.

Perspective

The demonstrated scope is tabular data: held-out random structural causal models, five hand-built graph motifs, ACIC-2016 real covariates with simulated outcomes, the Hillstrom e-mail marketing randomized trial augmented with constructed decoys, and the Causal Chambers light-tunnel physical device; the main model is trained on tables of up to 15 features, with an appendix extension to 64 features.

FADO learns the Bayesian posterior mean under its training prior, so how closely it matches the true causal profile depends on whether that profile is identifiable from observational tables; under linear-Gaussian and hidden-confounder generating processes, observational data are consistent with multiple causal structures and performance drops. The authors report reliable ranking and causal-versus-non-causal separation but not accurate point estimates of effect sizes; bootstrap confidence intervals are not calibrated, though their correlation with error can be used to screen out the most uncertain features.