Denoising Diffusion Causal Discovery
Keywords: Denoising Diffusion Models, Causal Discovery, Causal Reasoning, Diffusion Models, Graph Neural Networks
TL;DR: We use denoising diffusion models to do causal network inference by suggesting a denoising objective.
Abstract: A common theme across multiple disciplines of science is to understand the underlying dependencies between variables from observational data. Such dependencies are often modeled as Bayesian Network (BNs), which by definition are Directed Acyclic Graphs (DAGs). Recent advancements, such as NOTEARS and DAG-GNN, have focused on formulating continuous DAG constraints and learning DAGs via continuous optimization. However, these methods often have scalability issues and face challenges when applied to real world data. In this paper, we propose Denoising Diffusion Causal Discovery (DDCD), a new learning framework that leverages Denoising Diffusion Probabilistic Models (DDPMs) for causal structural learning. Using the denoising objective, our method allows the model to explore a wider range of noise in the data and effectively captures both linear and nonlinear dependencies. It also has reduced complexity and is more suitable for inference of larger networks. To accommodate potential feedback loops in biological networks, we propose a k-hop acyclicity constraint. Additionally, we suggest using fixed-size bootstrap sampling to ensure similar training performance across varying dataset sizes. Our experiments on synthetic data demonstrate that DDCD achieves consistent competitive performance compared to existing methods while noticeably reducing computation time. We also show that DDCD can generate trustworthy networks from real-world datasets.
Primary Area: causal reasoning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13285
Loading