DeDiCa: A Scalable Deconfounding Discrete Causal Generative Model

Key points

• Given a causal graph, DeDiCa is trained once on observational data by maximum likelihood, exactly marginalizing over hidden confounders, and can then estimate any (proxy-)identifiable interventional query.1
• Despite this deliberately simple parameterization, DeDiCa matches state-of-the-art neural causal generative models in ATE estimation, while supporting memory-efficient CPU training and exact interventional inference in milliseconds.1
• Compared with causal foundation models, DeDiCa scales more effectively with additional observational samples and achieves competitive accuracy at substantially lower inference time and memory.1
• DeDiCa provides a tractable, accurate, and scalable approach to causal inference in discrete settings with hidden confounding.1

Sources (1)

• [1]DeDiCa: A Scalable Deconfounding Discrete Causal Generative Model
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:42 AM

  We introduce DeDiCa, a practical framework for efficient causal inference with discrete variables and hidden confounding that revisits classical probabilistic modeling through a tensorized parameterization of causal Bayesian networks, enabling tractable inference.

  Given a causal graph, DeDiCa is trained once on observational data by maximum likelihood, exactly marginalizing over hidden confounders, and can then estimate any (proxy-)identifiable interventional query.

Extractive summary: sentences quoted from the sources.