Research map / Graph learning / Knowledge and temporal graphs
Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs
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 TMLP, which replaces GCN message passing in multimodal recommendation with MLPs trained via topological pruning of item-item graphs and a neighborhood alignment (MINE-based) loss, outperforming nine baselines on three Amazon datasets.
Problem
GCN-based multimodal recommenders suffer from over-smoothing, so they can only use shallow GCNs with limited representation power, and their item-item graphs built from pre-trained modality features are noisy due to modality incompatibility.
Approach
Bypass GCNs entirely: use MLPs for multimodal item encoding, inject item-item topology via a neighborhood alignment loss derived from mutual information maximization (MINE), and denoise the item graph with a topological pruning strategy based on neighborhood-overlap mutual information before training.
Key results
- TMLP achieves Recall@10 of 0.0671 on Baby, a 7.53% improvement over the second-best model (MGCN/FREEDOM family baselines), and improves Recall@10 by 5.49% on Sports and 5.69% on Electronics.
- Training efficiency: TMLP reaches peak Recall@20 by about epoch 30 on Sports and epoch 10 on Electronics, with total training cost of 10.2 min versus 32 min for FREEDOM and BM3 (per appendix table).
- Under 1% corrupted item-adjacency noise on Sports, TMLP maintains performance while FREEDOM and MGCN drop significantly, sometimes below LightGCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest baseline with TMLP | TMLP versus strongest baseline · Baby | R@10: 7.53% | Helped |
| replace the strongest baseline with TMLP | TMLP versus strongest baseline · Sports | R@10: 5.49% | Helped |
| replace the strongest baseline with TMLP | TMLP versus strongest baseline · Electronics | R@10: 5.69% | Helped |
| add visual features to textual features (ablation) | TMLP · Baby | R@20: 0.1010 -> 0.1016 | Helped |
| add visual features to textual features (ablation) | TMLP · Sports | R@20: 0.1128 -> 0.1152 | Helped |
| add textual features to visual features (ablation) | TMLP · Baby | R@20: 0.0847 -> 0.1016 | Helped |
| add textual features to visual features (ablation) | TMLP · Sports | R@20: 0.0981 -> 0.1152 | Helped |
| add topological pruning (ablation) | TMLP · Baby | R@20: 0.0926 -> 0.1016 | Helped |
| add topological pruning (ablation) | TMLP · Sports | R@20: 0.0838 -> 0.1152 | Helped |
| add neighborhood alignment loss (ablation) | TMLP · Baby | R@20: 0.0812 -> 0.1016 | Helped |
| add neighborhood alignment loss (ablation) | TMLP · Sports | R@20: 0.0948 -> 0.1152 | Helped |
| replace random pruning with topological pruning (ablation) | TMLP · Baby | R@20: 0.0993 -> 0.1016 | Helped |
Surprising findings
- Random pruning of item neighbors performs slightly better than no pruning in ablation, attributed to small K filtering out noise.
- Textual features contribute more to performance than visual features on Amazon datasets (TMLP_T closer to full TMLP than TMLP_V).
Limitations
- Gains are reported on three Amazon datasets with a single backbone aggregator (LightGCN), so generality is not shown beyond these settings.
- Hyperparameters such as K, alpha, and learning rate are tuned by grid search, and the reported best settings may be dataset-specific.
- Figure-level claims about over-smoothing are supported mainly by qualitative visualizations rather than formal analysis.
Experimental setup
Datasets: Amazon Baby Amazon Sports Amazon Electronics
Benchmarks: Amazon Baby Amazon Sports Amazon Electronics
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: BPR LightGCN VBPR MMGCN GRCN DualGNN BM3 MGCN FREEDOM
Builds on: FREEDOM LightGCN MINE BPR MMRec
Closest papers by meaning
- Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommendation (AAAI 2025 · similarity 0.92)
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.90)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.89)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
- GraphVLM: Benchmarking Vision Language Models for Multimodal Graph Learning (CVPR 2026 · similarity 0.88)
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.