ACTIVA: Amortized Causal Effect Estimation via Variational Autoencoders

Andreas W.M. Sauter, Saber Salehkaleybar, Aske Plaat, Erman Acar

Keywords: causal inference, variational autoencoder, amortized inference, zero-shot learning

TL;DR: Introduces ACTIVA, a transformer-based conditional VAE that enables zero-shot estimation of interventional distributions from observational data, amortizing across causal problems.

Abstract: Predicting the distribution of outcomes under hypothetical interventions is crucial in healthcare, economics, and policy-making. However, existing methods often require restrictive assumptions, and are typically limited by the lack of amortization across problem instances. We propose ACTIVA, a transformer-based conditional variational autoencoder (VAE) architecture for amortized causal inference, which estimates interventional distributions directly from observational data. ACTIVA learns a latent representation conditioned on observational inputs and intervention queries, enabling zero-shot inference by amortizing causal knowledge from diverse training scenarios. We provide theoretical insights showing that ACTIVA predicts interventional distributions as mixtures over observationally equivalent causal models. Empirical evaluations on synthetic and semi-synthetic datasets validate our insights and show the effectiveness of our amortized approach, highlighting promising directions for future real-world applications.

Primary Area: causal reasoning

Submission Number: 5708