Research map / Label-efficient and robust learning / Recognition and long-tail
Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec
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 FENRec, which adds time-dependent soft labels built from near-future interactions and enduring hard negatives mixed from anchors and in-batch negatives, integrated into ICSRec for sequential recommendation.
Problem
Data sparsity in sequential recommendation: contrastive methods use binary next-item labels that ignore upcoming interactions, and random negative sampling yields easy negatives in later training stages.
Approach
Generate soft labels over the next few future items (decaying with time via a smoothing parameter gamma) from existing subsequences, and create hard negatives by mixing the anchor with in-batch negatives throughout training (stop-gradient, norm-preserved), plus a focal-style hard-negative upweighting contrastive loss.
Key results
- FENRec achieves the best HIT@5 on all four datasets, e.g., 0.0431 vs. 0.0400 for the best baseline BSARec on Sports (7.75% gain).
- On Yelp, FENRec improves HIT@10 by 12.27% and NDCG@10 by 11.82% over the best baseline.
- Average improvement of 6.16% across all metrics is claimed in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FENRec instead of the strongest baseline | Sequential recommendation · Sports | HIT@5: 0.0400 -> 0.0431 | Helped |
| use FENRec instead of the strongest baseline | Sequential recommendation · Beauty | HIT@5: 0.0707 -> 0.0728 | Helped |
| use FENRec instead of the strongest baseline | Sequential recommendation · Toys | HIT@5: 0.0792 -> 0.0818 | Helped |
| use FENRec instead of the strongest baseline | Sequential recommendation · Yelp | HIT@5: 0.0260 -> 0.0286 | Helped |
| add Time-Dependent Soft Labeling (ablation) | FENRec · Sports | HIT@5: 0.0403 -> 0.0431 | Helped |
| add Enduring Hard Negatives Incorporation (ablation) | FENRec · Sports | HIT@5: 0.0422 -> 0.0431 | Helped |
| add Time-Dependent Soft Labeling (ablation) | FENRec · Beauty | HIT@5: 0.0693 -> 0.0728 | Helped |
| add Enduring Hard Negatives Incorporation (ablation) | FENRec · Beauty | HIT@5: 0.0707 -> 0.0728 | Helped |
| add Time-Dependent Soft Labeling (ablation) | FENRec · Toys | HIT@5: 0.0789 -> 0.0818 | Helped |
| add Enduring Hard Negatives Incorporation (ablation) | FENRec · Toys | HIT@5: 0.0806 -> 0.0818 | Helped |
| add Time-Dependent Soft Labeling (ablation) | FENRec · Yelp | HIT@5: 0.0275 -> 0.0286 | Helped |
| add Enduring Hard Negatives Incorporation (ablation) | FENRec · Yelp | HIT@5: 0.0268 -> 0.0286 | Helped |
Surprising findings
- Removing time-dependent soft labels hurts more than removing enduring hard negatives, per the ablation study.
Limitations
- Hard negatives mixing anchor with in-batch negatives can create false negatives, sensitive to batch size.
- Does not incorporate auxiliary time information (noted as future work).
Experimental setup
Datasets: Amazon Sports Amazon Beauty Amazon Toys Yelp
Benchmarks: Amazon Sports Amazon Beauty Amazon Toys Yelp
Metrics: HR@K NDCG@K
Baselines: GRU4Rec Caser LRURec SASRec BSARec BERT4Rec MAERec CBiT CL4SRec CoSeRec ICLRec DuoRec
Builds on: ICSRec focal-infoNCE DuoRec SASRec MVS
Closest papers by meaning
- Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.90)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.89)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.88)
- SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.88)
- LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.88)
- Intent Oriented Contrastive Learning for Sequential Recommendation (AAAI 2025 · 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.