Test-Time Learning of Causal Structure from Interventional Data

Wei Chen, Rui Ding, Huang Bojun, Yuxuan Liang, Yang Zhang, Qiang Fu, Shi Han, Dongmei Zhang

Keywords: Test-Time Traing, Self-Augmentation, Causal Structure Learning, Intervention Target Detection

TL;DR: We introduce a novel approach for causal discovery from interventional data that adapts to test distribution shifts using test-time training, improving accuracy by generating tailored training data.

Abstract:

Inferring causal structures from interventional data remains a challenging task, especially when the intervention targets are unknown. *Supervised Causal Learning (SCL)* demonstrates strong empirical performance in predicting causal structures by training on datasets with known causal relations and applying the learned models to unseen test data. However, existing *SCL* methods often face inherent generalization challenges and struggle with the diverse intervention settings encountered in the *interventional causal discovery* problem.

In this work, we propose _**TICL**_ (**T**est-time **I**nterventional **C**ausal **L**earning), a novel approach that follows the *Test-Time Training (TTT)* + *Joint Causal Inference (JCI)* paradigm to address these challenges of generalization and versatility, respectively. Specifically, _**TICL**_ employs a self-augmentation technique that generates training data at test time, tailored to the characteristics of the test data, enabling the model to adapt to the inherent biases in the test distribution. Additionally, by integrating the *JCI* framework with *SCL*, _**TICL**_ replaces the rule-based logic of the standard PC algorithm with a learning-based approach, effectively leveraging self-augmented training data.

Extensive experiments on bnlearn benchmarks demonstrate _**TICL**_'s superiority in multiple aspects of causal discovery and intervention target detection.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

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: 3081