Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Time Slot Preferences via Mobility Tree 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 the Mobility Tree, a hierarchical day/period/check-in structure, and MTNet, a Tree-LSTM-based multitask model that learns time-slot-specific user preferences for next POI recommendation.
Problem
Next POI recommendation ignores users' personalized preferences across distinct time slots within a day, as existing models use only sequential or graph structures over check-ins.
Approach
Encode a user's trajectory as a multi-granularity tree (day nodes, period nodes, check-in leaves), propagate information bottom-up with intra-hierarchy self-attention and inter-hierarchy N-ary Tree-LSTM in four steps, and train with an uncertainty-weighted multitask loss predicting next POI, geographic cluster, and category.
Key results
- MTNet achieves Acc@1 of 0.2620 on NYC, 0.2575 on TKY, and 0.1453 on CA, outperforming the strongest baseline GETNext (0.2435, 0.2254, 0.1357).
- On CA, MTNet improves over GETNext by 15.26% in Acc on average and 12.55% in MRR.
- Ablation on TKY: removing IAC drops Acc@1 from 0.2575 to 0.2353 and removing IRC drops it to 0.2391; removing multitask learning lowers MRR from 0.3659 to 0.3499.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MTNet instead of GETNext | Next POI recommendation · NYC | Acc@1: 0.2435 -> 0.2620 | Helped |
| use MTNet instead of GETNext | Next POI recommendation · TKY | Acc@1: 0.2254 -> 0.2575 | Helped |
| use MTNet instead of GETNext | Next POI recommendation · CA | Acc@1: 0.1357 -> 0.1453 | Helped |
| add multitask learning (ablation) | MTNet · TKY | Acc@1: 0.2410 -> 0.2575 | Helped |
| add geography prediction (ablation) | MTNet · TKY | Acc@1: 0.2353 -> 0.2575 | Helped |
| add category prediction (ablation) | MTNet · TKY | Acc@1: 0.2398 -> 0.2575 | Helped |
| add geography and category prediction (ablation) | MTNet · TKY | Acc@1: 0.2377 -> 0.2575 | Helped |
| add intra-hierarchy communication (IAC) (ablation) | MTNet · TKY | Acc@1: 0.2353 -> 0.2575 | Helped |
| add inter-hierarchy communication (IRC) (ablation) | MTNet · TKY | Acc@1: 0.2391 -> 0.2575 | Helped |
| include current period node and last check-in node predictions (ablation) | MTNet · TKY | Acc@1: 0.2478 -> 0.2575 | Helped |
Surprising findings
- Using very few (2 or 3) or very many (20 or 24) time slots degrades performance, so the optimal slot count (12 for TKY/CA) is tied to check-in timestamp density.
Limitations
- Only time-slot granularity and a fixed 24-hour trajectory split are explored; cross-day structure is not modeled beyond day nodes.
- Experiments are limited to three LBSN datasets and ten baselines, with no reported variance across runs.
Experimental setup
Datasets: Foursquare NYC Foursquare TKY Gowalla CA
Benchmarks: Foursquare NYC Foursquare TKY Gowalla CA
Metrics: Acc@1 Acc@5 Acc@10 MRR
Baselines: FPMC PRME LSTM ST-RNN STGN STGCN PLSPL CFPRec STAN GETNext
Builds on: Tree-LSTM Transformer multi-task uncertainty weighting
Closest papers by meaning
- Integrating Personalized Spatio-Temporal Clustering for Next POI Recommendation (AAAI 2025 · similarity 0.87)
- GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model (AAAI 2025 · similarity 0.86)
- 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 0.86)
- Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (AAAI 2025 · similarity 0.86)
- Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · 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.