Counterfactual prediction skips the causal graph: a GNN reads observational samples directly, and a 10-node model transfers to 30 nodes

Counterfactual prediction no longer has to estimate a causal graph first and then run inference on it; the counterfactual distribution comes straight from observational samples. The prior route estimated the causal graph from observational data and then used that graph for counterfactual inference, but the space of causal graphs grows super-exponentially with the number of variables, and choosing the wrong graph amplifies errors into the downstream counterfactual. On synthetic benchmarks including Triangle, M-shape and German Loan, CFGNP improves across almost all variants over VACA and BGM, which need the full causal graph, and over CFM, which needs only a topological ordering; on German Loan its MSE is lower than CFM's, while on the LIN variant of M-shape its MSE is higher than BGM's.

The model places observational samples, the factual observation and the counterfactual condition in one graph and passes messages so that information crosses between worlds only along corresponding factual and counterfactual variables. It uses neural processes to map a set of observations to a predictive distribution, then adds graph neural networks so the structural constraints of the counterfactual dual graph enter the prediction directly instead of a single estimated graph being fixed first. On CheXpert chest X-rays, CFGNP reduces MAE on age relative to DeepBC and reaches higher ROC-AUC on categorical variables such as disease than DeepBC and VCI; on synthetic graphs, a model trained only on 10-node graphs keeps strong predictive performance up to 30 nodes, though it is worse than per-size retraining under the Log-sum-exp and Sqrt-sum mechanisms.

The work demonstrates predicting counterfactual distributions directly from observational samples when the causal graph is unknown, evaluated on synthetic benchmarks including Triangle, M-shape and German Loan, on randomly generated causal graphs of increasing size, and on the CheXpert chest X-ray image benchmark. It is aimed at readers working on causal inference, graph neural networks and image-based counterfactual generation. The authors propose incorporating priors about the causal graph, such as knowledge of the presence or absence of certain causal relationships, or simple distributions over edges.

The authors note that the required observational context increases both training and inference costs, and that when the counterfactual distribution is non-identifiable CFGNP cannot recover the true distribution, which they state also applies to the baselines. On CheXpert, ground-truth counterfactuals are unavailable, so evaluation relies on classifiers judging whether generated images exhibit the intended attributes, and images are compared qualitatively.