FADO: Learning Causal Explanations with Prior-Fitted Networks
Organizations: University of Oxford · RPTU Kaiserslautern-Landau · Ellison Institute of Technology · Adelaide University
Abstract
In regression and classification tasks with tabular datasets, most of the literature on explanation in ML (e.g. SHAP, permutation importance) has focused on predictive attribution: does feature help predict outcome ? At its heart, predictive attribution captures correlation, not causation. We instead ask a causal question: would manipulating feature change outcome ? We define a natural quantity to answer this question, the feature intervention effect: . In general, estimating this quantity from observational data alone would entail strong causal assumptions. We propose FADO (Feature-wise Amortized Do-Operators), an amortized inference method using prior-fitted networks trained on synthetic tabular datasets generated by structural causal models. During training, the synthetic data-generating process is known, so we can compute ground-truth intervention effects , and train a model to predict them at test time from a purely observational tabular dataset. As we demonstrate through extensive experiments, our new causal explainability method is easy to train and provides better explanations under a range of settings for which standard predictive attribution fails, including proxy variables, confounders, colliders, and placebos. We evaluate it on real, synthetic and semi-synthetic datasets, with and without hidden confounders, and compare it to existing predictive attribution methods and state-of-the-art causal inference methods.
Figures & tables
| Quantity | Definition | What it answers |
|---|---|---|
| Marginal correlation | Is associated with ? | |
| Shapley value / SHAP | averaged over subsets | Does help a predictor of ? |
| Linear regression coefficient | from | Marginal linear contribution holding others fixed. |
| Permutation importance | when is permuted | Predictive importance of in a fitted model. |
| LiNGAM total effect | Identified effect under linear non-Gaussian SCMs. | |
| Causal feature profile | , see eq. 2 | Would intervening on change ? |
| FADO vs | Marginal baseline | |||
|---|---|---|---|---|
| Prior family | Pearson | AUROC (nonzero) | MAE (zero) | Pearson |
| Linear | ||||
| Mixed type | ||||
| Polynomial nonlinear | ||||
| Random MLP | ||||
| Hidden confounder | ||||
| Linear | Nonlinear | Hidden | Mixed | |||||
| Gaussian | Non-Gaussian | Polynomial | MLP | confounder | type | |||
| Causal discovery | ||||||||
| LiNGAM | ||||||||
| NOTEARS | ||||||||
| PC | ||||||||
| Noise | Proxy | Leak | Conf-child | |||||||||
| Family | F | Mg | Mv | F | Mg | Mv | F | Mg | Mv | F | Mg | Mv |
| No confounding | ||||||||||||
| Linear | — | — | — | |||||||||
| Mixed nonlinear | — | — | — | |||||||||
| Interaction | — | — | — | |||||||||
| Hidden confounder | ||||||||||||
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
| FADO vs | Marginal baseline | |||
|---|---|---|---|---|
| Training setting | Pearson | AUROC (nonzero) | MAE (zero) | Pearson |
| FADO | ||||
| Linear , 15k steps | ||||
| Linear , 20k steps | ||||
| Linear , 30k steps | ||||
| Gaussian control | ||||
| FADO vs | Marginal baseline | |||
|---|---|---|---|---|
| Training setting | Pearson | AUROC (nonzero) | MAE (zero) | Pearson |
| Linear , 15k steps | ||||
| Large ( , 4+4) | ||||
| Large ( , 4+4) | ||||
| Gaussian control | ||||
| MLP + hidden confounder | ||||
| Scale | Seed 0 | Seed 1 | Seed 2 | mean std |
|---|---|---|---|---|
| Eval | Heterogeneous | Dedicated | Gap |
|---|---|---|---|
| pts | |||
| pts | |||
| pts |
| FADO | do-Shapley † | Causal Shapley † | DSCM † | Marginal | Multivariate | |
|---|---|---|---|---|---|---|
| Linear | ||||||
| Nonlinear | ||||||
| MLP | ||||||
| Hidden confounder | ||||||
| MLP + hidden |
| Eval | Density | Signal frac. | Metric | FADO | FADO | FADO , fixed | Baseline |
|---|---|---|---|---|---|---|---|
| paper | Pearson | ||||||
| MSE | |||||||
| AUROC (nonzero) | |||||||
| paper | Pearson | ||||||
| MSE | |||||||
| AUROC (nonzero) |
| Variant | FADO | vs Marginal corr. | Marginal corr. | Multivariate regression |
|---|---|---|---|---|
| Linear , 15k | ||||
| , 15k | ||||
| , 20k | ||||
| Foundation |
| Configuration | Pearson | AUROC (nonzero) | Change in Pearson | |
|---|---|---|---|---|
| FADO (all components) | — | |||
| sample axis attention | ||||
| feature axis attention | ||||
| type embedding (feature/ ) |
| Type-id config | Pearson( ) | MAE (zero) | Pearson( ) |
|---|---|---|---|
| Normal (col marked) | |||
| All zero (no column marked) | |||
| Mispointed (col marked) |
| Motif | DAG | Expected |
|---|---|---|
| A: direct cause | ||
| B: observed confounder | ||
| C: mediator chain | ||
| D: target descendant | ||
| E: collider |
| Motif (structure) | Key feature | Truth | Model | Marginal corr. | Multivariate regression | SHAP |
|---|---|---|---|---|---|---|
| A: direct cause | ||||||
| B: confounded | ||||||
| C: mediator | / | |||||
| D: target descendant | ||||||
| E: collider |
| Empirical coverage (all features), nominal | Coverage@95 | ||||||
|---|---|---|---|---|---|---|---|
| Regime | Cause | Null | |||||
| Linear (identifiable) | |||||||
| Linear (non-ident.) | |||||||
| Hidden conf. (non-ident.) | |||||||
| MLP | |||||||
| Mixture | |||||||
| FADO bootstrap | ensemble | CausalPFN native | ||||
| Regime | sel. gain | sel. gain | sel. gain | |||
| Linear (ident.) | ||||||
| Linear (non-ident.) | — | — | ||||
| Hidden conf. | ||||||
| MLP | ||||||
| Mixture | — | — | — | — | — | — |
| Feature | FADO | SHAP | Marginal | Backdoor | |
|---|---|---|---|---|---|
| red | |||||
| green | |||||
| blue | |||||
| pol_1 | |||||
| pol_2 | |||||
| l_11 |
| Pearson | Proxy | Noise | |||||
|---|---|---|---|---|---|---|---|
| Config | FADO | Marginal corr. | Multivariate regression | FADO | Marginal corr. | FADO | Marginal corr. |
| diabetes / linear | |||||||
| diabetes / mixed nonlin. | |||||||
| california / linear | |||||||
| california / mixed nonlin. | |||||||
| abalone / interaction | |||||||
| Family | Weight | Generator | Label |
|---|---|---|---|
| Linear | Exact | ||
| Mixed type | linear; each feature independently median-split to with probability | Exact | |
| Polynomial nonlinear | MC | ||
| Random MLP | , a random one-hidden-layer MLP | MC | |
| Hidden confounder | linear with one latent parent of three observed nodes, marginalized at sample time | Exact from full DAG | |
| MLP hidden confounder | random MLP with the same latent structure | MC from full DAG |
| Tag | Prior family | Steps | Batch | |
| Linear | linear non-Gaussian | |||
| Linear , 50k | linear non-Gaussian | |||
| Linear | linear non-Gaussian | |||
| Linear | linear non-Gaussian | |||
| Linear | linear non-Gaussian | |||
| Large | linear non-Gaussian |
| Family | Raw draw | Random parameters |
|---|---|---|
| laplace | — | |
| gaussian | — | |
| student_t3 | — | |
| lognormal_signed | , sign-flipped w.p. | |
| exp_signed | , Rademacher | — |
| gauss_mixture_bimodal |
| Setting | CausalPFN ‡ | Do-PFN ‡ | FADO |
|---|---|---|---|
| Linear non-Gaussian | |||
| Linear non-Gaussian | |||
| Linear non-Gaussian | |||
| Linear Gaussian | |||
| Polynomial nonlinear | |||
| Random MLP |
| Linear | Nonlinear | Hidden | Mixed | |||||
| Gauss. | Non-Gaussian | Poly. | MLP | conf. | type | |||
| Causal discovery | ||||||||
| LiNGAM | ||||||||
| NOTEARS | ||||||||
| PC | ||||||||
| , | , | ||||
|---|---|---|---|---|---|
| Method | Cost type | CPU (s) | GPU (s) | CPU (s) | GPU (s) |
| Marginal corr. | inference | — | — | ||
| Multivariate regression | inference | — | — | ||
| LightGBM-SHAP | fit explain | — | — | ||
| DoubleML (LinearDML) | fit explain | — | — | ||
| Causal Forest | fit explain | — | — | ||
| Feature | SHAP rank | DR-Learner | FADO | Truth | |
|---|---|---|---|---|---|
| Leak | #1 | ||||
| Proxy | #7 | ||||
| noise 1 | — | ||||
| noise 2 | — | ||||
| noise 3 | — | ||||
| #3 | RCT |
| Dataset | Audits | Flagged | |
|---|---|---|---|
| Hillstrom | |||
| LaLonde | |||
| Criteo | ( f2 ) | ||
| Sachs | |||
| Tübingen | varies | ( pair0024 , pair0094 ) | |
| Total | — |
| Estimator | MSE if | MSE if | Worst-case MSE |
|---|---|---|---|
| Posterior mean | 0.360 | 0.360 | 0.360 |
| Marginal association | 0.001 | 1.442 | 1.442 |
| Zero effect | 1.440 | 0.000 | 1.440 |
| No-free-lunch lower bound | 0.360 |