ACTIVA: Amortized Causal Effect Estimation via Variational Autoencoders

Andreas Sauter [email protected], Saber Salehkaleybar [email protected], Frank van Harmelen [email protected], Aske Plaat [email protected], Erman Acar [email protected] · 2025

Read ACTIVA: Amortized Causal Effect Estimation via Variational Autoencoders by Andreas Sauter [email protected], Saber Salehkaleybar [email protected], Frank van Harmelen [email protected], Aske Plaat [email protected], Erman Acar [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-01290.txt.

Description

Andreas Sauter and co-authors present a method to predict what happens when a system is intervened on, in a 2025 paper. Such predictions usually need either experiments or a known causal map, and each new task often needs fresh fitting. Their system, called ACTIVA, learns from many training tasks at once using a transformer-based model.

Once trained, the system answers a new task in a single pass, without intervention data or a causal graph. The authors prove what the system aims at when data and capacity are ideal: a blend of the outcomes from all causal models that fit the observations. On synthetic and gene-based tests, it beat a correlation-based rival and matched strong baselines. It also cut false effects on unrelated variables.

Catalog record

CREATORAndreas Sauter [email protected], Saber Salehkaleybar [email protected], Frank van Harmelen [email protected], Aske Plaat [email protected], Erman Acar [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE100 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-01290

Browse the full library · About the Tower of Babel