Research map / Label-efficient and robust learning / Recognition and long-tail
BeFA: A General Behavior-driven Feature Adapter for Multimedia 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 BeFA, a plug-and-play behavior-driven adapter that decouples, gates and reconstructs pre-extracted multimodal content features using behavioral embeddings, and an attribution analysis showing information drift and omission in those features.
Problem
Pre-trained feature encoders extract content features from entire items, including preference-irrelevant details, so the features may not reflect user preferences in multimedia recommendation.
Approach
A similarity-based (cosine to behavioral embeddings) CAM-style attribution method reveals information drift and omission in content features; BeFA then adapts content features via a linear decoupling followed by behavior-gated filtering and recombination, trained end-to-end with the recommender.
Key results
- Adding BeFA to BM3, LATTICE, FREEDOM and MGCN improves Recall@20 and NDCG@20 on average by 9.07% and 11.02% across datasets and encoders (per paper text).
- On H&M, BeFA increases Recall@20 and NDCG@20 by 14.17% and 17.34%; on Microlens by 4.58% and 5.19%.
- BeFA adds about 0.20M parameters (0.93% to 2.09% of the system) and raises per-epoch training time by roughly 15%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BeFA to MGCN | MGCN with CLIP features · TMALL | Recall@20: 0.0380 -> 0.0395 | Helped |
| Add BeFA to MGCN | MGCN with CLIP features · Microlens | Recall@20: 0.0972 -> 0.1000 | Helped |
| Add BeFA to MGCN | MGCN with CLIP features · H&M | NDCG@20: 0.0251 -> 0.0274 | Helped |
| Add BeFA to BM3 | BM3 with CLIP features · Microlens | Recall@20: 0.0851 -> 0.0911 | Helped |
| Add BeFA to BM3 | BM3 with CLIP features · TMALL | Recall@20: 0.0298 -> 0.0319; BM3+LoRA: 0.0333 | Mixed |
| Use a small decoupling space (0.125x or 0.25x) (ablation) | BeFA · All datasets | Recall@20 | Hurt |
| Vary the decoupling-space size (ablation) | BeFA · Microlens | Recall@20 | Mixed |
Surprising findings
- Even the more advanced ImageBind encoder still yields content features with information drift and omission, according to the attribution analysis.
Limitations
- Decoupling space size is a hyperparameter; performance fluctuates considerably across its values and datasets.
- Results vary across datasets; gains on Microlens are smaller than on H&M.
Experimental setup
Datasets: TMALL Microlens H&M
Benchmarks: TMALL Microlens H&M
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: LoRA Soft Prompt Tuning LATTICE BM3 FREEDOM MGCN
Builds on: LATTICE BM3 FREEDOM MGCN Class Activation Mapping
Closest papers by meaning
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.84)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.83)
- Multiple Purchase Chains with Negative Transfer Elimination for Multi-Behavior Recommendation (AAAI 2025 · similarity 0.83)
- MTRec: Learning to Align with User Preferences via Mental Reward Models (NeurIPS 2025 · similarity 0.83)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.83)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.83)
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.