Research map / Label-efficient and robust learning / Recognition and long-tail
Fine Tuning Out-of-Vocabulary Item Recommendation with User Sequence Imagination
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 USIM, an RL-based framework that fine-tunes generated OOV item embeddings by imagining user sequences and backpropagating through imagined users, using a recommendation-aware reward and RecPPO training.
Problem
Out-of-vocabulary (OOV) item recommendation, where new items lack behavioral interaction embeddings and only content features are available, so 'makeshift' embeddings generated from content suffer a content-behavior gap and suboptimal recommendation performance.
Approach
Instead of only generating a one-shot embedding from content features, imagine a sequence of users who interacted with the OOV item and refine its embedding via simulated backpropagation (embedding-driven state transition), with the sequence selection cast as an MDP solved by a recommender-oriented PPO using embedding-alignment and recommendation-performance rewards.
Key results
- On CiteULike with MF, USIM reaches overall Recall 0.1926 vs. 0.1584 for the best baseline ALDI (about 19% improvement), with statistical significance p<0.01.
- In a two-week online A/B test on a major e-commerce platform (5% of users per group), USIM improved OOV item GMV by 20.30% over Random, 14.95% over MetaEmb, and 13.60% over ALDI.
- On MovieLens with GNN backbone, USIM OOV Recall 0.1448 vs. 0.1429 for ALDI, a modest gain; ablation shows removing the performance reward lowers CiteULike MF overall Recall from 0.1926 to 0.1866.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune OOV item embeddings by imagining user sequences | USIM with MF · CiteULike OOV | Recall: 0.2622 -> 0.2753 | Helped |
| fine-tune OOV item embeddings by imagining user sequences | USIM with MF · MovieLens OOV | NDCG: 0.1445 -> 0.1466 | Helped |
| fine-tune OOV item embeddings by imagining user sequences | USIM with GNN · CiteULike OOV | NDCG: 0.1399 -> 0.1452 | Helped |
| fine-tune OOV item embeddings by imagining user sequences | USIM with GNN · MovieLens OOV | NDCG: 0.1567 -> 0.1638 | Helped |
| add cosine-similarity top-k user selection to the positive action set (ablation) | USIM · CiteULike OOV, MF and GNN | Recall: 0.2336 -> 0.2753; 0.2363 -> 0.2534 | Helped |
| add randomly sampled actions to the exploration set (ablation) | USIM · CiteULike OOV, MF and GNN | Recall: 0.2341 -> 0.2753; 0.2411 -> 0.2534 | Helped |
| construct an exploration set instead of exploring the entire action space (ablation) | USIM · CiteULike OOV, MF and GNN | Recall: 0.2232 -> 0.2753; 0.1971 -> 0.2534 | Helped |
| add recommendation-performance reward (ablation) | USIM · CiteULike OOV, MF and GNN | Recall: 0.2711 -> 0.2753; 0.2414 -> 0.2534 | Helped |
| use the full USIM imagination strategy rather than randomly selecting users | USIM · CiteULike OOV and overall | NDCG | Helped |
| randomly sample a user at each imagination step | user-sequence optimization · CiteULike OOV and overall | NDCG | Hurt |
Surprising findings
- Dropout-based baselines (e.g., Heater, CLCRec) degrade IV item recommendation, while generative methods and USIM preserve it.
- Randomly sampling users at each imagination step causes a sharp, non-recoverable drop in performance, and performance declines as more users are imagined beyond an optimum.
Limitations
- Many hyperparameters (k, N, p, lambda) require costly grid search tuning.
- Autoregressive-style sequential generation of imagined users is computationally intensive and slow.
- Evaluation limited to two benchmark datasets plus one industrial A/B test; industrial results lack detailed methodology in main text.
Experimental setup
Datasets: CiteULike MovieLens
Benchmarks: CiteULike MovieLens
Metrics: Recall@K NDCG@K OOV item PV OOV item PCTR OOV item GMV
Baselines: DropoutNet MTPR Heater CLCRec DeepMusic MetaEmb GAR ALDI Wolpertinger Policy Hierarchical Reinforcement Learning
Builds on: PPO DropoutNet ALDI GAR MetaEmb
Closest papers by meaning
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.88)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.87)
- Language Representations Can be What Recommenders Need: Findings and Potentials (ICLR 2025 · similarity 0.85)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.85)
- LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.85)
- On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · similarity 0.85)
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.