Research map / Graph learning / Knowledge and temporal graphs
Successive POI Recommendation via Brain-Inspired Spatiotemporal Aware Representation
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 STEP, a brain-inspired (entorhinal-hippocampal) unsupervised spatiotemporal embedding framework (grid-cell spatial encoder, visiting-time encoder, spatiotemporal context graph) feeding RNN recommenders for successive POI recommendation.
Problem
Successive POI recommendation suffers from sub-optimal POI embeddings because spatial and temporal natures are modeled separately and POI-specific spatiotemporal patterns are underused.
Approach
Builds context graphs (sequential, spatial, spatiotemporal) and learns POI embeddings unsupervised via negative-sampling graph objectives; a grid-cell multi-scale spatial encoder and a POI visiting-time pattern encoder are fused conjunctively into a spatiotemporal representation, concatenated with sequential embeddings and fed to a simple RNN recommender.
Key results
- STEP+LSTM reaches HIT@1 0.2454 and MRR 0.2835 on Instagram Check-in, versus 0.1622 HIT@1 and 0.2093 MRR for the best two-stage baseline (Geo+GRU, MRR 0.2093 on Gowalla comparison row).
- On Gowalla, STEP+LSTM achieves HIT@10 0.3728 and MRR 0.2282, outperforming STP-UDGAT (HIT@10 0.3285, MRR 0.2130).
- Ablation shows removing the spatiotemporal model (keeping only sequential) drops Gowalla MRR from 0.2282 to 0.1683, with larger gains from temporal information on the richer-timestamp Instagram dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STEP+LSTM instead of Geo+GRU | Successive POI recommendation · Instagram Check-in | HIT@1: 0.1619 -> 0.2454 | Helped |
| Use STEP+LSTM instead of Geo+GRU | Successive POI recommendation · Instagram Check-in | MRR: 0.2093 -> 0.2835 | Helped |
| Use STEP+LSTM instead of STP-UDGAT | Successive POI recommendation · Gowalla | HIT@1: 0.1475 -> 0.1539 | Helped |
| Use STEP+LSTM instead of STP-UDGAT | Successive POI recommendation · Gowalla | MRR: 0.2130 -> 0.2282 | Helped |
| Add POI-specific temporal information processing (ablation) | STEP+LSTM · Instagram Check-in and Gowalla | HIT@1: 0.2433 -> 0.2454; 0.1461 -> 0.1539 | Helped |
| Add the visiting-time encoder (ablation) | STEP+LSTM · Instagram Check-in and Gowalla | MRR: 0.2798 -> 0.2835; 0.2248 -> 0.2282 | Helped |
| Use the grid-cell encoder instead of a one-layer location encoder (ablation) | STEP+LSTM · Instagram Check-in and Gowalla | HIT@1: 0.2399 -> 0.2454; 0.1006 -> 0.1539 | Helped |
| Add spatial embedding preserving (ablation) | STEP+LSTM · Instagram Check-in and Gowalla | HIT@1: 0.2298 -> 0.2454; 0.0973 -> 0.1539 | Helped |
| Remove the sequential model and use only the spatiotemporal model (ablation) | STEP+LSTM · Instagram Check-in | HIT@1: 0.2454 -> 0.2466 | Mixed |
| Remove the sequential model and use only the spatiotemporal model (ablation) | STEP+LSTM · Gowalla | HIT@1: 0.1539 -> 0.1252 | Hurt |
Surprising findings
- Removing the sequential model improves HIT@1 on Instagram (0.2466 vs 0.2454), suggesting coarse sequential information can hurt top-1 precision.
Limitations
- Spatiotemporal context graph sparsity depends on hyperparameters h and m that must be tuned per dataset.
- Relies on spatial and temporal context graphs built with simple unweighted edges and fixed grid-scale ranges.
Experimental setup
Datasets: Instagram Check-in Gowalla TaxiBJ
Benchmarks: Instagram Check-in Gowalla TaxiBJ
Metrics: HIT@1 HIT@5 HIT@10 MRR
Baselines: Random+GRU Skip-Gram+GRU CAPE Geo+GRU ST-RNN STGN LSTPM STP-UDGAT
Builds on: Geo (grid-cell POI embedding) skip-gram word2vec negative sampling
Closest papers by meaning
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.87)
- Learning Place Cell Representations and Context-Dependent Remapping (NeurIPS 2024 · similarity 0.86)
- Time Makes Space: Emergence of Place Fields in Networks Encoding Temporally Continuous Sensory Experiences (NeurIPS 2024 · similarity 0.84)
- Flexible Context-Driven Sensory Processing in Dynamical Vision Models (NeurIPS 2024 · similarity 0.84)
- A simple model of co-emergence of grid and place fields (NeurIPS 2026 · similarity 0.84)
- REMI: Reconstructing Episodic Memory During Internally Driven Path Planning (NeurIPS 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.