Research map / Large language models / Language, safety and interpretability
POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation 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 POI-Enhancer, a portable framework that extracts POI semantic features from a frozen Llama-2-7B via three prompts and fuses them into classic POI embeddings (Skip-Gram, POI2Vec, Geo-Teaser, TALE, Hier, CTLE) using alignment, cross-attention, and multi-view contrastive learning.
Problem
Existing POI representations use weak textual features (categories or check-in text), and it is unclear how to extract task-general POI knowledge from LLMs and integrate it into POI embeddings.
Approach
Three role-play prompts (visit pattern, address, surrounding) elicit LLM last-hidden-state features, which are aligned (dual feature alignment via Transformer encoders), fused by attention weighting, and injected into existing POI embeddings through multi-query cross-attention, trained with temporal/spatial/functional multi-view contrastive learning.
Key results
- Enhanced baselines (+) improve over originals on all three datasets and tasks; e.g., Hier+ Hit@5 on NY rises from 15.631 to 19.197 (+22.81%).
- On check-in sequence classification, CTLE+ on SG gains 23.05% Acc and 39.62% F1 over CTLE and gains are largest in this task (e.g., POI2Vec+ on NY +62.68% Acc).
- Removing the LLM features (Hier+ w/o LLM) drops NY POI recommendation Hit@1 from 8.009 to 7.029.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Enhance POI representations with POI-Enhancer | Six POI embedding baselines · Three downstream tasks across NY, TKY, and SG | Task-specific metrics | Helped |
| Add LLM semantic features (ablation) | Hier+ · POI Recommendation, NY | Hit@1: 7.029 -> 8.009 | Helped |
| Add LLM semantic features (ablation) | Hier+ · POI Recommendation, NY | Hit@5: 17.028 -> 19.197 | Helped |
| Keep the framework's fusion process and contrastive learning instead of replacing fusion with vector addition and removing contrastive learning (ablation) | Hier+ · POI Recommendation, NY | Hit@1: 8.009 -> 8.009 | No clear effect |
| Keep the framework's fusion process and contrastive learning instead of replacing fusion with vector addition and removing contrastive learning (ablation) | Hier+ · POI Recommendation, NY | Hit@5: 19.197 -> 19.197 | No clear effect |
| Use specialized prompts with role-playing, attribute headers, and questions (ablation) | POI-Enhancer · New York downstream tasks | Hit@1, Acc, and MAE | Helped |
| Use Dual Feature Alignment and Semantic Feature Fusion (ablation) | POI-Enhancer · New York downstream tasks | Hit@1, Acc, and MAE | Helped |
| Use Cross Attention Fusion instead of concatenating POI and LLM vectors (ablation) | POI-Enhancer · New York downstream tasks | Hit@1, Acc, and MAE | Helped |
| Use temporal, spatial, and functional multi-view contrastive sampling instead of spatial-only sampling (ablation) | POI-Enhancer · New York downstream tasks | Hit@1, Acc, and MAE | Helped |
| Increase alignment and fusion layer counts past their optimal settings (ablation) | POI-Enhancer with Hier · POI Recommendation, NY | Hit@1 | Hurt |
Surprising findings
- Raw LLM features (GPT-2, Llama2, ChatGLM2) alone perform far worse than the enhanced POI embedding models on all tasks, indicating LLM embeddings alone are insufficient for POI tasks.
Limitations
- Relies on a single LLM backbone (Llama-2-7B) with frozen features; generalization to other LLMs not studied in the main text.
- Some improvements are modest and results appear to come from a single split without reported variance.
- Address and surrounding attributes depend on external Nominatim API queries and heuristic preprocessing.
Experimental setup
Datasets: Foursquare-NY Foursquare-SG Foursquare-TKY
Benchmarks: Foursquare-NY Foursquare-TKY Foursquare-SG
Metrics: Hit@1 Hit@5 ACC macro-F1 MAE RMSE NMI
Baselines: Skip-Gram POI2Vec Geo-Teaser TALE Hier CTLE GPT-2 Llama2 ChatGLM2
Builds on: InfoNCE LibCity Skip-Gram CTLE
Closest papers by meaning
- Mixture of In-Context Experts Enhance LLMs' Long Context Awareness (NeurIPS 2024 · similarity 0.82)
- Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection (CVPR 2024 · similarity 0.82)
- LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs (ICML 2026 · similarity 0.82)
- MM1.5: Methods, Analysis & Insights from Multimodal LLM Fine-tuning (ICLR 2025 · similarity 0.81)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.81)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.81)
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.