Modeling Causal Mechanisms with Diffusion Models for Interventional and Counterfactual Queries
Patrick Chao University of Pennsylvania Work done during internship at Amazon. Patrick Blöbaum Amazon Sapan Patel Amazon Shiva Prasad Kasiviswanathan Amazon
Abstract
We consider the problem of answering observational, interventional, and counterfactual queries in a causally sufficient setting where only observational data and the causal graph are available. Utilizing the recent developments in diffusion models, we introduce diffusion-based causal models (DCM) to learn causal mechanisms, that generate unique latent encodings. These encodings enable us to directly sample under interventions and perform abduction for counterfactuals. Diffusion models are a natural fit here, since they can encode each node to a latent representation that acts as a proxy for exogenous noise. Our empirical evaluations demonstrate significant improvements over existing state-of-the-art methods for answering causal queries. Furthermore, we provide theoretical results that offer a methodology for analyzing counterfactual estimation in general encoder-decoder models, which could be useful in settings beyond our proposed approach.
中文速览
扩散模型天生具备"编码潜变量"的能力，研究者借此提出了一种用于回答因果查询的新框架——基于扩散的因果模型（Diffusion-based Causal Model, DCM）：针对因果图中的每个节点单独训练一个扩散模型，利用DDIM的确定性正向/反向过程将观测值编码为潜在表示，以此作为结构因果模型中外生噪声的代理，从而统一支持观测、干预和反事实三类因果查询。在合成数据集（涵盖多种结构方程类型）和真实fMRI数据上的实验表明，DCM相比现有最优方法（CAREFL、VACA等）有显著提升，对于非加性噪声模型下的干预和反事实查询甚至领先一个数量级以上，训练速度也快数倍。此外，论文还给出了编码-解码框架下反事实估计误差的首个理论界，该理论不局限于扩散模型，对其他编码-解码类生成模型同样适用，为因果推断领域提供了更坚实的理论基础。
原文 arXiv:2302.00860;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2302.00860v3