ARTFEED — Contemporary Art Intelligence

FePySR: Neural Feature Extraction for Symbolic Regression

other · 2026-05-14

FePySR is a two-stage framework for symbolic regression that first uses a heterogeneous neural network to extract valid features from observational data, then performs structural optimization with PySR. It addresses the NP-hard problem of recovering complex mathematical expressions by decomposing them into reusable nonlinear feature modules. On five standard benchmarks, FePySR achieves higher equation recovery rates than state-of-the-art methods. On 75 highly complex synthesized equations, it recovers 36 equations with substantially smaller mean squared error.

Key facts

  • FePySR is a two-stage framework for symbolic regression.
  • It uses a heterogeneous neural network for feature extraction.
  • Structural optimization is performed using PySR.
  • It reduces the SR search space by extracting valid features prior to equation search.
  • Outperforms state-of-the-art methods on five standard benchmarks.
  • Recovers 36 out of 75 highly complex synthesized equations.
  • Produces substantially smaller mean squared error.
  • Addresses NP-hard problem of recovering complex mathematical expressions.

Entities

Sources