Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning
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 FHCRec, which pairs a frequency-enhanced bidirectional Transformer for long-term check-in sequences with local and global hierarchical contrastive learning over short-term subsequences for next POI recommendation.
Problem
Next POI recommendation where Transformer-based long-term sequence encoders suffer over-smoothing (losing high-frequency information), and short-term sequences are defined only by the most recent check-ins and modeled independently from long-term sequences.
Approach
Applies FFT-based low- and high-frequency enhancement in parallel with self-attention inside a bidirectional Transformer for long-term sequences, and treats date-segmented subsequences within the long-term sequence as short-term patterns refined via local and global contrastive losses.
Key results
- FHCRec reaches HR@5 0.6000 and NDCG@5 0.4257 on PHO, versus 0.5368 and 0.3811 for the strongest baseline CLSPRec, an improvement of roughly 7% to 21% across datasets.
- On SIN, FHCRec achieves HR@10 0.4219... reported improvements of 19% to 21% in NDCG over CLSPRec on SIN.
- Ablation shows removing the frequency-enhanced encoder drops PHO HR@5 from 0.6000 to 0.5579.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the frequency-enhanced Transformer instead of a standard encoder (ablation) | FHCRec · PHO, NYC, SIN | HR@5: 0.5579 -> 0.6000; 0.3856 -> 0.3917; 0.4051 -> 0.4219 | Helped |
| Add the frequency-enhanced Transformer instead of a standard encoder (ablation) | FHCRec · PHO, NYC, SIN | NDCG@10: 0.4454 -> 0.4481; 0.2937 -> 0.3099; 0.3561 -> 0.3562 | Helped |
| Add global contrastive learning (ablation) | FHCRec · PHO, NYC, SIN | HR@5: 0.5737 -> 0.6000; 0.3784 -> 0.3917; 0.4145 -> 0.4219 | Helped |
| Add global contrastive learning (ablation) | FHCRec · PHO, NYC, SIN | NDCG@10: 0.4428 -> 0.4481; 0.3061 -> 0.3099; 0.3552 -> 0.3562 | Helped |
| Add local contrastive learning (ablation) | FHCRec · PHO, NYC, SIN | HR@5: 0.5789 -> 0.6000; 0.3773 -> 0.3917; 0.4119 -> 0.4219 | Helped |
| Add local contrastive learning (ablation) | FHCRec · PHO, NYC, SIN | NDCG@10: 0.4421 -> 0.4481; 0.3068 -> 0.3099; 0.3539 -> 0.3562 | Helped |
| Add the contrastive learning module (ablation) | FHCRec · PHO | HR@5: 0.5526 -> 0.6000 | Helped |
| Add the contrastive learning module (ablation) | FHCRec · PHO | NDCG@10: 0.4411 -> 0.4481 | Helped |
Surprising findings
- Removing the frequency-enhanced encoder lowers HR@5 on PHO but barely changes NDCG@10 on NYC and SIN, suggesting gains vary by metric and dataset.
Limitations
- Frequency threshold hyperparameter d and contrast weight alpha are tuned per dataset on test-set-style sweeps, with results reported under a grid search.
- Only three Foursquare datasets are evaluated, all from check-in data.
- Transformer-based long-term modeling still relies on the bidirectional Transformer backbone.
Experimental setup
Datasets: Foursquare Singapore (SIN) Foursquare New York City (NYC) Foursquare Phoenix (PHO)
Benchmarks: Foursquare SIN Foursquare NYC Foursquare PHO
Metrics: HR@5 HR@10 NDCG@5 NDCG@10
Baselines: ST-RNN ATST-LSTM MCARNN PLSPL iMTL CTLE CFPRec ContraRec CLSPRec
Builds on: CLSPRec CFPRec ContraRec
Closest papers by meaning
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.87)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.87)
- GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model (AAAI 2025 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.86)
- Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation (AAAI 2024 · 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.