Research map / Video / Video segmentation and tracking
Integrating Personalized Spatio-Temporal Clustering for Next POI Recommendation
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 iPCM, a next-POI recommendation model that clusters POIs spatially (k-means) and time slices temporally (ordered sample clustering), fuses POI/user/region/time embeddings into a Transformer encoder with MLP heads, and adds a probability adjustment step using user-region-POI and user-period-POI preferences.
Problem
Next POI recommendation on LBSN check-in data suffers from broad search scopes across whole cities and all times, and individual user data is too sparse to learn personalized spatio-temporal clustering effects.
Approach
Learn personalized spatio-temporal clustering by clustering POIs into k-means regions and day time slices into ordered-sample periods, using global trajectory data plus user embeddings, then predict next region and period to narrow the POI search via a probability adjustment step.
Key results
- On the NYC dataset, the full iPCM reaches Acc@1 0.2604, Acc@10 0.66 and MRR 0.3962, versus 0.2236 Acc@1 for the POI-only baseline.
- Across all four datasets, iPCM is reported to improve on average 452.96% over MF, 49.13% over FPMC, 8.29% over GETNext and 5.88% over AGRAN.
- On TKY, iPCM reaches Acc@1 0.4044 and MRR 0.5136, improving over POI+user (0.3691 Acc@1, 0.488 MRR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add user information to the POI-only model (ablation) | iPCM · NYC | Acc@1: 0.2236 -> 0.2486 | Helped |
| Add user information to the POI-only model (ablation) | iPCM · TKY | Acc@1: 0.2556 -> 0.3691 | Helped |
| Add personalized temporal clustering to POI+user (ablation) | iPCM · NYC | Acc@1: 0.2486 -> 0.2604 | Helped |
| Add personalized temporal clustering to POI+user (ablation) | iPCM · TKY | Acc@1: 0.3691 -> 0.4044 | Helped |
| Add personalized spatial clustering to POI+user (ablation) | iPCM · NYC | Acc@1: 0.2486 -> 0.2604 | Helped |
| Add personalized spatial clustering to POI+user (ablation) | iPCM · TKY | Acc@1: 0.3691 -> 0.4044 | Helped |
| Add both personalized temporal and spatial clustering to POI+user (ablation) | iPCM · NYC | Acc@1: 0.2486 -> 0.2604 | Helped |
| Add both personalized temporal and spatial clustering to POI+user (ablation) | iPCM · TKY | Acc@1: 0.3691 -> 0.4044 | Helped |
| Add personalized temporal clustering to POI+user (ablation) | iPCM · NYC | Acc@5: 0.553 -> 0.5456 | No clear effect |
| Add personalized temporal clustering to POI+user (ablation) | iPCM · TKY | Acc@10: 0.7149 -> 0.7151 | No clear effect |
| Add personalized spatial clustering to POI+user (ablation) | iPCM · NYC | Acc@5: 0.553 -> 0.55 | No clear effect |
| Add personalized spatial clustering to POI+user (ablation) | iPCM · TKY | Acc@5: 0.6358 -> 0.6524 | Helped |
Surprising findings
- Directly adjusting POI probabilities with region predictions did not improve results, because regions contain many POIs and introduce noise.
Limitations
- Evaluated only on Foursquare datasets (NYC, TKY, SIN, OSK) with a small set of baselines.
- Spatial and temporal cluster numbers are hyperparameters; sensitivity results show little variation but no principled selection is given.
- Baseline comparisons are summarized as average percentage improvements rather than per-dataset tables in the main text.
Experimental setup
Datasets: Foursquare NYC Foursquare TKY Foursquare SIN Foursquare OSK
Benchmarks: Foursquare NYC Foursquare TKY Foursquare SIN Foursquare OSK
Metrics: Acc@1 Acc@5 Acc@10 MRR
Baselines: MF FPMC GETNext AGRAN
Builds on: GETNext STAN GGNN Time2Vec
Closest papers by meaning
- GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model (AAAI 2025 · similarity 0.90)
- Learning Time Slot Preferences via Mobility Tree for Next POI Recommendation (AAAI 2024 · similarity 0.87)
- TAU: Trajectory Data Augmentation with Uncertainty for Next POI Recommendation (AAAI 2024 · similarity 0.86)
- SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation (NeurIPS 2025 · similarity 0.84)
- Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (AAAI 2025 · similarity 0.84)
- Spatial-Temporal Knowledge Distillation for Takeaway Recommendation (AAAI 2025 · similarity 0.84)
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.