Research map / Label-efficient and robust learning / Recognition and long-tail
Top-one Recommendation with Anonymous User Behaviors (Student Abstract)
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 RRLC, a training objective combining a reciprocal rank loss (with flexibility factor sigma) and a listwise contrastive loss over frequent-pattern-augmented sessions, to improve session-based recommendation on top-one MRR and HR.
Problem
Session-based recommendation (top-one recommendation with anonymous user behavior) must optimize MRR, which is hard to optimize with gradient descent due to vanishing gradients at bottom ranks, and anonymous sessions are very short, making augmentation-based contrastive learning unreliable.
Approach
A reciprocal rank loss 1/(r_s-1)^sigma that approximates exact reciprocal rank with an adjustable gradient at bottom ranks, plus a listwise contrastive loss where sessions are augmented via frequent-pattern insertion and consistency is measured by a differentiable approximation of rank-biased overlap.
Key results
- RRLC improves NARM, SRGNN and ATTMIX across Diginetica and Tmall, with average gains of 7.38% HR@5 and 6.78% MRR@5 reported in the text.
- On Tmall, SRGNN with RRLC improves HR@5 by 13.48% and MRR@5 by 12.56% over baseline.
- Ablation: removing the reciprocal rank loss (replacing with BCE) or the listwise contrastive loss generally lowers performance compared to full RRLC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize NARM with RRLC | NARM · Diginetica | MRR@5: 14.97 -> 15.88 | Helped |
| Optimize NARM with RRLC | NARM · Tmall | MRR@5: 13.89 -> 14.97 | Helped |
| Optimize SRGNN with RRLC | SRGNN · Diginetica | MRR@5: 15.26 -> 16.17 | Helped |
| Optimize SRGNN with RRLC | SRGNN · Tmall | MRR@5: 13.12 -> 14.77 | Helped |
| Optimize ATTMIX with RRLC | ATTMIX · Diginetica | MRR@5: 15.52 -> 15.99 | Helped |
| Optimize ATTMIX with RRLC | ATTMIX · Tmall | MRR@5: 17.34 -> 18.23 | Helped |
| Add reciprocal rank loss instead of BCE (ablation) | NARM + RRLC · Diginetica | MRR@5: 15.48 -> 15.88 | Helped |
| Add listwise contrastive loss (ablation) | NARM + RRLC · Diginetica | MRR@5: 15.73 -> 15.88 | Helped |
| Add reciprocal rank loss instead of BCE (ablation) | NARM + RRLC · Tmall | MRR@5: 13.97 -> 14.97 | Helped |
| Add listwise contrastive loss (ablation) | NARM + RRLC · Tmall | MRR@5: 13.99 -> 14.97 | Helped |
| Add reciprocal rank loss instead of BCE (ablation) | SRGNN + RRLC · Diginetica | MRR@5: 16.00 -> 16.17 | Helped |
| Add listwise contrastive loss (ablation) | SRGNN + RRLC · Diginetica | MRR@5: 15.83 -> 16.17 | Helped |
Limitations
- Ablation gains are small and not consistent across all settings (e.g., w/o LC sometimes matches or exceeds full RRLC on ATTMIX).
- Evaluation limited to two datasets and three SBR baselines, all reported at short cutoff K=5.
- Gradient behavior and sigma choice are shown only via illustrative curves rather than extensive tuning analysis.
Experimental setup
Datasets: Diginetica Tmall
Benchmarks: Diginetica Tmall
Metrics: HR@5 MRR@5
Baselines: NARM SR-GNN AttMix
Builds on: NARM SR-GNN AttMix
Closest papers by meaning
- Dual Sparse Attention Network For Session-based Recommendation (AAAI 2021 · similarity 0.86)
- Bid Farewell to Seesaw: Towards Accurate Long-Tail Session-Based Recommendation via Dual Constraints of Hybrid Intents (AAAI 2026 · similarity 0.86)
- Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.85)
- ScoreNet: Consistency-driven Framework with Multi-side Information Fusion for Session-based Recommendation (AAAI 2025 · similarity 0.84)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.83)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · 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.