NBSE: Physics-Informed Feature Selection Eliminates Greedy Search via Nishimori Temperature
arXiv·low signal
Usatyuk, Sapozhnikov, and Egorov propose Noise-Based Spectral Embedding, a framework that selects informative features from high-dimensional data by constructing a sparse similarity graph and identifying the Nishimori temperature — the critical point where the Bethe Hessian becomes singular. The eigenvector at this critical temperature captures the dominant diffusion mode, naturally reweighting nodes and eliminating computationally expensive greedy feature search.