The Path-Driven Independence Testing (PIT) Algorithm
Keywords: Causality, Bayesian networks, Structure learning, Constraint based, PC algorithm
Abstract: PC is an efficient constraint-based algorithm for learning the structure of a Bayesian network. However, the required number of conditional independent (CI) tests can make the algorithm practically infeasible or slow for large graphs. We developed a constrained-based algorithm, called the Path-Driven Independence Testing (PIT) Algorithm, which during the learning process, utilizes the information of the partially learned network to reduce the number of CI tests. The idea is that for each pair of variables $X$ and $Y$, instead of checking independence conditioned on every subset of all the neighbors of $X$ (resp. $Y$) as in PC, the search is restricted to only the common neighbors of $X$ and $Y$ and to neighbors connected to $Y$ (resp. $X$) by a path. Also, paths connecting $X$ and $Y$ without a descendant of a common neighbor can be blocked by observing two consecutive nodes on the path. Compared to PC, PIT is proven to conduct at most the same number of CI tests, and experimentally shown to be significantly (up to 7 times) faster and more accurate.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
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: 7561
Loading