Go to ICLR 2025 Conference homepageGeoBench: A new benchmark on Symbolic Regression with Geometric Expressions
Keywords: Symbolic Regression, Geometry
Abstract: Symbolic regression (SR) is a powerful technique for deriving mathematical expressions from data. With the emergence of numerous SR methods, SRBench made a significant contribution by providing a standardized testing platform that includes 130 SR datasets and evaluates 14 SR methods. However, the methods included in SRBench are outdated, and the dataset does not feature results from more recent approaches such as TPSR. Additionally, the metrics used in SRBench do not adequately capture the full capabilities of symbolic regression methods, and the benchmark data has scientific problems. Although Matsubara et al. (2022) address some of these issues,
their approach remains incomplete. In response, we propose a new benchmark consisting of 71 expressions derived from geometric contexts, categorized into three difficulty levels: easy, medium, and hard. We evaluate 20 SR methods on these expressions, focusing exclusively on the symbolic regression capability of each model, assessed through recovery rates across the different levels and overall. We provide a detailed methodology for reproducing the experiments and include results for newly developed SR methods within this updated benchmark. The results demonstrate significant variation in symbolic regression ability across models.
Primary Area: datasets and benchmarks
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: 6782
Loading