Research map / Vision-language and multimodal / Cross-modal learning
MENTOR: Multi-level Self-supervised Learning 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 MENTOR, a multimodal recommender that uses GCN and item-item graphs per modality plus multilevel Gaussian-distribution cross-modal alignment guided by ID embeddings and two general feature enhancement SSL tasks (feature masking and graph perturbation).
Problem
Multimodal recommendation suffers from label sparsity and from modality noise when aligning heterogeneous visual and textual features, and prior SSL alignment methods erode historical user-item interaction information.
Approach
Align visual, textual, and fused modality representations at four levels, using the ID embedding as guidance and modeling each representation as a Gaussian (mean and std matching), while adding feature-masking and graph-perturbation contrastive tasks for robustness; ID embeddings are used only for alignment, not for final prediction.
Key results
- On Baby, Sports, and Clothing, MENTOR improves Recall@20 over the best multimodal baseline by 5.64%, 4.59%, and 5.10% (per the Improv. row of the table).
- Recall@20 of MENTOR reaches 0.1048 on Baby, 0.1139 on Sports, and 0.0989 on Clothing, versus 0.0992, 0.1089, 0.0941 for FREEDOM.
- Ablation shows removing the general feature enhancement component (MENTOR_fg) drops Recall@20 to 0.1011, 0.1094, 0.0949 on Baby, Sports, Clothing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MENTOR instead of FREEDOM | multimodal recommendation · Baby | Recall@20: 0.0992 -> 0.1048 | Helped |
| use MENTOR instead of FREEDOM | multimodal recommendation · Sports | Recall@20: 0.1089 -> 0.1139 | Helped |
| use MENTOR instead of FREEDOM | multimodal recommendation · Clothing | Recall@20: 0.0941 -> 0.0989 | Helped |
| add general feature enhancement (ablation) | MENTOR · Baby, Sports, Clothing | Recall@20: 0.1011 -> 0.1048; 0.1094 -> 0.1139; 0.0949 -> 0.0989 | Helped |
| add feature masking task (ablation) | MENTOR · Baby, Sports, Clothing | Recall@20: 0.1037 -> 0.1048; 0.1126 -> 0.1139; 0.0981 -> 0.0989 | Helped |
| add graph perturbation task (ablation) | MENTOR · Baby, Sports, Clothing | Recall@20: 0.1034 -> 0.1048; 0.1129 -> 0.1139; 0.0977 -> 0.0989 | Helped |
| add each successive level of multilevel cross-modal alignment (ablation) | MENTOR · recommendation datasets (figure; values not provided) | recommendation performance | Helped |
Surprising findings
- The paper states MENTOR improves the best multimodal baseline by 24.34% and 74.73% on Sports and Clothing in the text, inconsistent with the table's Improv. row (4.59% and 5.10%).
Limitations
- Text-based comparison claims of improvement are inconsistent with the reported table values.
- Evaluated only on three Amazon datasets with visual and textual modalities; ID, visual, textual only.
- Relies on hyperparameter grid search with dataset-specific optimal settings, and the item-item graph is frozen after initialization.
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: MF-BPR LightGCN LayerGCN VBPR MMGCN DualGNN LATTICE FREEDOM SLMRec BM3 MMSSL
Builds on: LightGCN BM3 FREEDOM LATTICE PPMDR SLMRec
Closest papers by meaning
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.90)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.90)
- Seeing Beyond Noise: Joint Graph Structure Evaluation and Denoising for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.88)
- Harnessing Multimodal Large Language Models for Multimodal Sequential Recommendation (AAAI 2025 · similarity 0.87)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.87)
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.