Research map / Vision-language and multimodal / Cross-modal learning
STAIR: Manipulating Collaborative and Multimodal Information for E-Commerce 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 STAIR, a multimodal collaborative filtering model that initializes item embeddings from whitened text/visual features and uses forward and backward stepwise graph convolutions to keep collaborative and multimodal information co-existing in e-commerce recommendation.
Problem
Multimodal recommendation in non-content-driven e-commerce, where raw textual/visual features are weakly predictive of user behavior, and where combining them with collaborative graph convolution causes modality erasure and modality forgetting.
Approach
Initialize item embeddings with whitened multimodal features and apply a dimension-dependent (stepwise) LightGCN in which early dimensions get collaborative smoothing and later dimensions keep multimodal content; then a backward stepwise convolution on gradient updates, guided by a modality kNN graph, counteracts forgetting.
Key results
- STAIR reaches Recall@20 of 0.1042 on Baby, 0.1117 on Sports and 0.0663 on Electronics, the best among compared methods.
- STAIR improves over LightGCN by roughly 10-24% and over FREEDOM by roughly 2-6% across metrics on the reported datasets.
- STAIR's training time and GPU memory are close to LightGCN (e.g., 1.45s/epoch and 490MB on Baby), far below MMSSL (27.70s/epoch, 3,048MB on Baby).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STAIR instead of FREEDOM | E-commerce recommendation · Baby | Recall@10: 0.0649 -> 0.0674 | Helped |
| Use STAIR instead of MMSSL | E-commerce recommendation · Baby | NDCG@20: 0.0442 -> 0.0453 | Helped |
| Use STAIR instead of FREEDOM | E-commerce recommendation · Sports | Recall@10: 0.0715 -> 0.0743 | Helped |
| Use STAIR instead of MMSSL | E-commerce recommendation · Sports | NDCG@20: 0.0482 -> 0.0503 | Helped |
| Use STAIR instead of FREEDOM | E-commerce recommendation · Electronics | Recall@10: 0.0427 -> 0.0440 | Helped |
| Use STAIR instead of FREEDOM | E-commerce recommendation · Electronics | NDCG@20: 0.0295 -> 0.0302 | Helped |
| Add modality initialization (ablation) | STAIR · Baby, Sports, and Electronics | Recommendation performance | Helped |
| Add forward stepwise convolution (FSC) (ablation) | STAIR · Baby | Recommendation performance | Helped |
| Add forward stepwise convolution (FSC) (ablation) | STAIR · Sports | Recommendation performance | No clear effect |
| Add backward stepwise convolution (BSC) (ablation) | STAIR · Baby, Sports, and Electronics | Recommendation performance | Helped |
| Use a slightly lower gamma (ablation) | STAIR · Baby | Recommendation performance | Helped |
| Emphasize textual rather than visual modality (ablation) | STAIR · Baby | Recommendation performance | Helped |
Surprising findings
- Behavior uncertainty under raw textual or visual features is close to that of random Gaussian noise, suggesting e-commerce interactions are rarely driven by content modalities.
- MMGCN, which performs graph convolution per modality, underperforms the non-multimodal MF-BPR baseline on these e-commerce datasets.
Limitations
- May not fully exploit raw multimodal features in content-driven scenarios such as news and video recommendation.
- Gains over FREEDOM are modest (2-6%), and the hyperparameters (gamma, k_m, weight decay) require per-dataset tuning.
Experimental setup
Datasets: Baby Sports Electronics
Benchmarks: Baby Sports Electronics
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20 training time per epoch GPU memory
Baselines: MF-BPR LightGCN JGCF MMGCN LATTICE BM3 FREEDOM MMSSL
Builds on: LightGCN FREEDOM LATTICE PageRank-inspired GCN
Closest papers by meaning
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.88)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.88)
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.87)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.86)
- MTSTRec: Multimodal Time-Aligned Shared Token Recommender (ICML 2025 · similarity 0.86)
- Modality-Independent Graph Neural Networks with Global Transformers 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.