Research map / Vision-language and multimodal / Cross-modal learning

Seeing Beyond Noise: Joint Graph Structure Evaluation and Denoising for Multimodal Recommendation

AAAI 2025 · Vision-language and multimodal › Cross-modal learning · approach: Unimodal-to-multimodal representation transfer · 8 citations (OpenAlex)

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 EVEN, a denoising framework for multimodal recommendation that denoises item-item semantic graphs via behavior-driven co-occurrence and prunes noisy user-item interactions via implicit gradient-based contribution scores, with self-guided structure learning for cross-modal alignment.

Problem

Multimodal recommender systems are sensitive to noise from semantic priors in item-item graphs built from multimodal features and noisy user-item interactions (mis-operations, exploration, account sharing), which degrade recommendation accuracy.

Approach

Jointly denoise both graphs without user-behavior assumptions: adjust the multimodal item-item similarity graph with a behavior co-occurrence matrix, and prune user-item edges with probabilities derived from item gradient-based impact in the previous training round, then align ID and multimodal embeddings via contrastive self-guided structure learning.

Key results

• EVEN achieves R@10 of 0.0667 on Baby, 0.0759 on Sports, and 0.0662 on Clothing, outperforming LGMRec and FREEDOM.
• Abstract reports an average accuracy improvement of 8.95% over LGMRec and 5.90% over FREEDOM without extending total training time.
• Ablation shows removing multimodal alignment and fusion (w/o MAF) causes the largest drop, e.g., R@10 falls from 0.0667 to 0.0511 on Baby.

Measured effects reported in the paper

Surprising findings

• FREEDOM-ADT, a traditional denoising method adapted to multimodal recommendation, underperforms the FREEDOM backbone it extends.

Limitations

• Per-epoch training time is higher than several baselines (3.46 s vs 1.78 s for FREEDOM).
• Evaluated only on three Amazon datasets with a single backbone family.
• Hyperparameters such as k1, k2, dropout ratio and layer counts require grid search and are sensitive.

Experimental setup

Datasets: Amazon Baby Amazon Sports and Outdoors Amazon Clothing, Shoes and Jewelry

Benchmarks: Amazon Baby Amazon Sports and Outdoors Amazon Clothing, Shoes and Jewelry

Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20

Baselines: VBPR LightGCN SLMRec LATTICE BM3 FREEDOM MMSSL LGMRec LayerGCN GRCN ADT DualGNN

Builds on: LATTICE FREEDOM LGMRec BPR LightGCN

Closest papers by meaning

• Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.90)
• Structured Spectral Reasoning for Frequency-Adaptive Multimodal Recommendation (NeurIPS 2025 · similarity 0.89)
• MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
• DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
• Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.87)
• LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.87)