Research map / Label-efficient and robust learning / Recognition and long-tail
SymSpectra: Symmetric Information Bottleneck Framework for Molecular Structure Recognition under Imbalanced Settings
Xiaohan Qin, Wenjie Du, Yang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SymSpectra's SCIB + CMI weighting + LSTM decoder instead of the best baseline | SymSpectra · Simulated spectra (IR, 13C-NMR, 1H-NMR) | F1: 0.947 -> 0.970 | Improved |
| add focal loss to the Alberts et al. baseline | Alberts et al. · Simulated spectra | F1: 0.947 -> 0.956 | Improved |
| use CMI dynamic weighting instead of focal loss, oversampling or reweighting | SymSpectra · Simulated spectra | F1: 0.956 -> 0.970 | Improved |
| add CMI dynamic weighting (ablation) | SymSpectra · low-frequency functional groups | F1: 0.844 -> 0.866 | Improved |
| add SCIB compression (ablation) | SymSpectra · low-frequency functional groups | F1: 0.827 -> 0.866 | Improved |
| add LSTM decoder for label dependencies (ablation) | SymSpectra · low-frequency functional groups | F1: 0.854 -> 0.866 | Improved |
| add LSTM decoder for label dependencies (ablation) | SymSpectra · high-frequency functional groups | F1: 0.966 -> 0.967 | No clear effect |
| use sequential prediction with IUPAC order instead of non-sequential CNN (ablation) | SymSpectra decoder · Imine functional group | F1: 0.540 -> 0.885 | Improved |
| use sequential prediction instead of non-sequential CNN (ablation) | SymSpectra decoder · Carboxylic acid functional group | F1: 0.992 -> 0.974 | Hurt |
| use scaffold-disjoint test split | SymSpectra · structurally disjoint test set | F1: 0.970 -> 0.8798 | Hurt |
| evaluate on experimental SDBS spectra instead of simulated | SymSpectra · multimodal spectra | F1: 0.970 -> 0.919 | Hurt |
| set s1 = s2 = 0.3 and beta = 1e-6 (ablation) | SymSpectra · hyperparameter sensitivity | F1 | Improved |
Closest papers by meaning
- Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.86)
- MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.86)
- Atom-Level Optical Chemical Structure Recognition with Limited Supervision (CVPR 2024 · similarity 0.85)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.85)
- MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.84)
- ChemPile: A 250 GB Diverse and Curated Dataset for Chemical Foundation Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
