Spectral Analysis of Molecular Features: When Richer Features Do Not Guarantee Better Generalization
A new paper analyzes the spectral properties of molecular features to understand model generalization in machine learning. Researchers found that richer spectral features do not always lead to better performance, challenging common assumptions in self-supervised learning. The study used kernel ridge regression across various representations like ECFP, transformers, and graph neural networks on QM9 and MoleculeNet benchmarks, revealing that only ECFP-based kernels showed a consistent positive correlation with performance. AI