Research map / Probabilistic and scientific modeling / Molecules and proteins
GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model
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 GeoMamba, a state-space-model POI recommender that combines a graph-based state space encoder (GaPPO) with an enhanced Mamba encoder and Matryoshka multi-scale representation learning to predict next POI visits.
Problem
Next POI recommendation from check-in sequences, where existing sequential models ignore continuous geographical/intention state transitions and multi-granular time-span preferences.
Approach
Extends HiPPO-based state space models to graph-structured geographical signals via a graph-autoregressive polynomial projection operator (GaPPO) computed in a reduced Graph Fourier spectral domain, combines it with a feature-enhanced Mamba encoder, and uses Matryoshka representation learning over multiple observation points to produce multi-scale user representations.
Key results
- GeoMamba achieves Recall@2 of 0.3840 on Singapore vs 0.3311 for the best baseline (Diff-POI), a 16.0% gain.
- On Tokyo, GeoMamba reaches Recall@2 0.6329 vs 0.5862 for Diff-POI (+7.97%).
- On New York City, GeoMamba reaches NDCG@2 0.6846 vs 0.6411 for Diff-POI (+6.79%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GeoMamba instead of Diff-POI | next-POI recommendation · Singapore | Recall@2: 0.3311 -> 0.3840⋆ | Helped |
| use GeoMamba instead of Diff-POI | next-POI recommendation · Tokyo | NDCG@2: 0.5748 -> 0.6224⋆ | Helped |
| use GeoMamba instead of Diff-POI | next-POI recommendation · New York City | Recall@2: 0.6501 -> 0.6893⋆ | Helped |
| use GeoMamba instead of Diff-POI | next-POI recommendation · New York City | NDCG@2: 0.6411 -> 0.6846⋆ | Helped |
| use specialized encoders instead of vanilla transformer layers (ablation) | GeoMamba · Singapore | Recall@2: 0.2719 -> 0.3740 | Helped |
| use specialized encoders instead of vanilla transformer layers (ablation) | GeoMamba · Tokyo | Recall@2: 0.5684 -> 0.6329 | Helped |
| use specialized encoders instead of vanilla transformer layers (ablation) | GeoMamba · New York City | Recall@2: 0.5667 -> 0.6893 | Helped |
| use specialized encoders instead of vanilla Mamba layers (ablation) | GeoMamba · Singapore | Recall@2: 0.2833 -> 0.3740 | Helped |
| use specialized encoders instead of vanilla Mamba layers (ablation) | GeoMamba · Tokyo | Recall@2: 0.5453 -> 0.6329 | Helped |
| use specialized encoders instead of vanilla Mamba layers (ablation) | GeoMamba · New York City | Recall@2: 0.6067 -> 0.6893 | Helped |
| add the enhanced Mamba encoder (ablation) | GeoMamba · Singapore | Recall@2: 0.3095 -> 0.3740 | Helped |
| add the enhanced Mamba encoder (ablation) | GeoMamba · Tokyo | Recall@2: 0.5636 -> 0.6329 | Helped |
Surprising findings
- Vanilla Transformer and vanilla Mamba layers perform similarly on POI recommendation and are both inferior to specialized recommendation models.
- Larger GaPPO filter window sizes do not necessarily yield better performance.
Limitations
- Ablation and sensitivity analyses are limited to three Foursquare datasets and a single fixed learning rate of 0.01.
- Performance gains decrease as SSM depth grows and eventually decline, indicating diminishing returns for deeper encoders.
Experimental setup
Datasets: Foursquare Singapore Foursquare Tokyo Foursquare New York City
Benchmarks: Foursquare Singapore Foursquare Tokyo Foursquare New York City
Metrics: Recall@K NDCG@K
Baselines: LightGCN HCCF GeoIE SGRec DRAN LSTPM STAN Diff-POI
Builds on: HiPPO Mamba Matryoshka representation learning graph ARMA filters Diff-POI
Closest papers by meaning
- Integrating Personalized Spatio-Temporal Clustering for Next POI Recommendation (AAAI 2025 · similarity 0.90)
- TAU: Trajectory Data Augmentation with Uncertainty for Next POI Recommendation (AAAI 2024 · similarity 0.87)
- Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (AAAI 2025 · similarity 0.86)
- Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation (AAAI 2024 · similarity 0.86)
- Spatial-Temporal Knowledge Distillation for Takeaway Recommendation (AAAI 2025 · similarity 0.85)
- SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation (NeurIPS 2025 · 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.