Efficient Subgraph Rule Induction via Tree Folding in Differentiable Logic Programming
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, 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.
Keywords: inductive logic programming, subgraph rules, gradient-based
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: This paper extends differentiable backwards chaining inductive logic programming techniques to support subgraph-shaped rules.
Abstract: Differentiable inductive logic programming techniques have proven effective at learning logic rules from noisy datasets; however, existing algorithms incur pernicious trade-offs between rule expressivity and scalability to large problems. Forward-chaining ILP algorithms can learn arbitrary rules, but their memory requirements scale exponentially with problem size. Backwards-chaining ILP algorithms address this limitation but do so with loss of generality by imposing the restrictive constraint that rules must be expressible as ensembles of independent chain-like Horn clauses. In this paper we present FUSE-ILP, a technique that relaxes this chain-like constraint and enables the differentiable evaluation of a restricted class of subgraph-like rules. Our method extends TensorLog-inspired backwards-chaining ILP techniques with branch masking and leaf grouping, which enable tree-like rule evaluation and “folding” of these trees into subgraphs. We demonstrate that this formulation allows our algorithm to learn more expressive rules than previous backwards-chaining algorithms while retaining a similar computational cost.
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: 8424
Loading