Research map / Vision-language and multimodal / Cross-modal learning
Structured Spectral Reasoning for Frequency-Adaptive Multimodal Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SSR, a four-stage spectral framework (decompose, modulate via spectral band masking, fuse via low-rank graph hyperspectral operator, align via spectral contrastive regularization) for frequency-adaptive multimodal graph recommendation, evaluated on three Amazon datasets.
Problem
Multimodal recommendation suffers from modality-specific noise, semantic inconsistency and unstable graph propagation; existing frequency-domain methods use static filtering or reweighting and lack band-level reasoning and reliability adaptation, especially under sparse and cold-start conditions.
Approach
Treats multimodal user-item graph signals as spectral bands, randomly masks bands during training with a representation-consistency loss, fuses bands with a low-rank hyperspectral operator (G-HSNO) that models cross-band and cross-modal interactions, and aligns modalities within each band via contrastive regularization.
Key results
- On Amazon Baby, SSR reports Recall@10 0.0728 vs 0.0680 for the strongest baseline SMORE.
- On Clothing, SSR improves Recall@10 by about 5% over SMORE (0.0708 vs 0.0659 as reported in the table).
- In the cold-start evaluation (users with at most 5 interactions), SSR outperforms all baselines on all twelve metrics, e.g. Baby R@20 0.1125 vs 0.1019 for SMORE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSR instead of SMORE | Multimodal recommendation · Baby | Recall@10: 0.0680 -> 0.0728 | Helped |
| Use SSR instead of AlignRec | Multimodal recommendation · Baby | Recall@20: 0.1046 -> 0.1103 | Helped |
| Use SSR instead of SMORE | Multimodal recommendation · Sports | Recall@10: 0.0762 -> 0.0825 | Helped |
| Use SSR instead of AlignRec | Multimodal recommendation · Sports | NDCG@10: 0.0414 -> 0.0449 | Helped |
| Use SSR instead of SMORE | Multimodal recommendation · Clothing | Recall@10: 0.0659 -> 0.0708 | Helped |
| Use SSR instead of AlignRec | Multimodal recommendation · Clothing | Recall@20: 0.0993 -> 0.1032 | Helped |
| Add semantic-aware fusion (SAF) (ablation) | SSR · Amazon recommendation | Performance | Helped |
| Add G-HSNO cross-band and cross-modality reasoning (ablation) | SSR · Amazon recommendation | Accuracy | Helped |
| Add Spectral Band Masking (SBM) (ablation) | SSR · Amazon recommendation | Robustness | Helped |
| Add Spectral Contrastive Regularization (SCR) (ablation) | SSR · Amazon recommendation | Alignment | Helped |
| Add multimodal inputs instead of using ID features only (ablation) | SSR · Amazon recommendation | Performance | Helped |
Limitations
- Spectral decomposition adds computational cost compared with time- or topology-domain approaches, per the authors.
- Number of frequency bands M is heuristic and may need dataset-specific tuning.
- Assumes a fixed transformation basis and focuses only on static multimodal graphs.
Experimental setup
Datasets: Amazon Baby Amazon Sports and Outdoors Amazon Clothing, Shoes and Jewelry
Benchmarks: Amazon Baby Amazon Sports Amazon Clothing
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: BPR LightGCN VBPR MMGCN GRCN DualGNN RecVAE LATTICE BM3 FREEDOM DiffMM MMIL
Builds on: SMORE LightGCN ChebyCF
Closest papers by meaning
- Seeing Beyond Noise: Joint Graph Structure Evaluation and Denoising for Multimodal Recommendation (AAAI 2025 · similarity 0.89)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.87)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.87)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.86)
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.86)
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.86)
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.