Research map / Graph learning / GNN architectures
Modality-Independent Graph Neural Networks with Global Transformers for 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 MIG-GT, which uses separate GNNs with modality-specific receptive fields (hops) for embedding, text and visual features plus a sampling-based global transformer with uniform global sampling, for multimodal recommendation.
Problem
Multimodal recommendation with GNNs on user-item graphs, where existing methods use the same receptive field (number of hops) for all modalities and limited hops restrict capture of global information.
Approach
Assigns independent GNN receptive fields K per modality (embedding, text, visual) using MGDN propagation, and adds a Sampling-based Global Transformer that attends each vertex to a small set of uniformly sampled global vertices, avoiding full quadratic attention, with an unsmooth regularization term.
Key results
- MIG-GT reaches R@20 0.1021 and N@20 0.0452 on Baby, versus 0.0992 R@20 for the strongest baseline FREEDOM, a roughly 5% improvement on Baby and Sports.
- On Sports MIG-GT achieves R@20 0.1130 and N@20 0.0511, and on Clothing R@20 0.0934, roughly matching FREEDOM (0.0932).
- MIG-GT outperforms graph transformer variants MIG-SGFormer and MIG-Polynormer, and with a contrastive loss (MIG-GT-CL) it is best on Clothing among compared CL methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Baby | R@20: 0.0992 -> 0.1021 | Helped |
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Baby | N@20: 0.0424 -> 0.0452 | Helped |
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Sports | R@20: 0.1089 -> 0.1130 | Helped |
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Sports | N@20: 0.0481 -> 0.0511 | Helped |
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Clothing | R@20: 0.0932 -> 0.0934 | Helped |
| use MIG-GT instead of FREEDOM | Multimodal recommendation · Clothing | N@20: 0.0416 -> 0.0422 | Helped |
| add Sampling-based Global Transformer (SGT) to MIRF (ablation) | MIG · Baby, Sports, and Clothing | NDCG@20 | Helped |
| replace SGT with SGFormer (ablation) | MIG-GT · Baby, Sports, and Clothing | R@20: 0.0863 vs 0.1021; 0.0887 vs 0.1130; 0.0827 vs 0.0934 | Hurt |
| replace SGT with Polynormer (ablation) | MIG-GT · Baby, Sports, and Clothing | R@20: 0.0997 vs 0.1021; 0.1048 vs 0.1130; 0.0864 vs 0.0934 | Hurt |
| add InfoNCE contrastive loss (ablation) | MIG-GT · Baby, Sports, and Clothing | R@20: 0.1021 -> 0.1022; 0.1130 -> 0.1120; 0.0934 -> 0.0946 | Mixed |
| add InfoNCE contrastive loss (ablation) | MIG-GT · Baby, Sports, and Clothing | N@20: 0.0452 -> 0.0451; 0.0511 -> 0.0505; 0.0422 -> 0.0428 | Mixed |
| increase global sample count from 5 to 10 (ablation) | SGT · Baby, Sports, and Clothing | performance | Helped |
Surprising findings
- Optimal receptive field K for some modalities on some datasets can be as low as 1 or 2, and differs across modalities (e.g., embedding and text best at K=3, visual at K=2 on Baby).
- The model does not use denoising or explicit item-item graphs, yet matches or outperforms those SOTA models.
Limitations
- Hyperparameter search (receptive fields, gamma, learning rate, L2) is required per dataset via grid search on validation set.
- Evaluated only on three Amazon datasets with pre-extracted text and visual features.
Experimental setup
Datasets: Amazon Baby Amazon Sports Amazon Clothing
Benchmarks: Amazon Baby Amazon Sports Amazon Clothing
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: MF LightGCN ApeGNN MGDN VBPR MMGCN GRCN DualGNN SLMRec LATTICE FREEDOM MMSSL
Builds on: MGDN LightGCN SGFormer Polynormer BPR
Closest papers by meaning
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.92)
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.90)
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.87)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.87)
- Mind Individual Information! Principal Graph Learning for Multimedia Recommendation (AAAI 2025 · similarity 0.86)
- Curriculum Conditioned Diffusion 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.