Research map / Label-efficient and robust learning / Recognition and long-tail
Cross-City Latent Space Alignment for Consistency Region Embedding
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 CoRE, a one-stage method that jointly learns mobility-based region embeddings for two cities while aligning their latent spaces via relative-representation manifold alignment and cross-city attention correlation consistency, enabling label-free cross-city transfer of socioeconomic predictors.
Problem
Region embeddings learned separately per city lie in incomparable latent spaces, so task predictors trained in a label-rich city cannot be reused in another city without labels; prior two-stage translation methods rely on hand-crafted cross-city correspondences and frozen embeddings.
Approach
Jointly train per-city GAT region embeddings with two alignment losses: a latent manifold alignment loss that compares similarities to shared virtual anchors (relative representations), and a latent individual alignment loss that uses cross-city attention to reconstruct intra-city pairwise region correlations, requiring no hand-crafted correspondences.
Key results
- CoRE reduces GDP MAE on XA->CD to 178.93 vs 183.68 for the best baseline HSA, with overall improvements of about 2.6% to 6.7% on GDP, population and carbon (paired t-test p<0.01).
- Ablations show removing either the latent manifold alignment or latent individual alignment loss degrades performance, and the one-stage design outperforms a two-stage variant.
- Parameter analysis finds alpha=beta=1 and about 1000 anchors work best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD GDP prediction · XA→CD GDP | MAE: 183.68 -> 178.93 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD GDP prediction · XA→CD GDP | MAPE: 11.51 -> 11.4 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD Population prediction · XA→CD Population | MAE: 642.17 -> 620.15 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD Population prediction · XA→CD Population | MAPE: 3.61 -> 3.42 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD Carbon prediction · XA→CD Carbon | MAE: 133.82 -> 127.45 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | XA→CD Carbon prediction · XA→CD Carbon | MAPE: 9.66 -> 9.42 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA GDP prediction · CD→XA GDP | MAE: 182.52 -> 170.25 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA GDP prediction · CD→XA GDP | MAPE: 5.6 -> 5.06 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA Population prediction · CD→XA Population | MAE: 614.32 -> 580.98 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA Population prediction · CD→XA Population | MAPE: 5.27 -> 4.88 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA Carbon prediction · CD→XA Carbon | MAE: 141.06 -> 136.56 | Helped |
| Use CoRE cross-city alignment instead of the strongest baseline | CD→XA Carbon prediction · CD→XA Carbon | MAPE: 1.97 -> 1.77 | Helped |
Limitations
- Evaluated only on three Chinese cities with one downstream predictor (ridge regression)
- Performance sensitive to number of anchors and loss weights
Experimental setup
Datasets: Xi'an mobility and socioeconomic data Chengdu mobility and socioeconomic data Beijing mobility and socioeconomic data
Benchmarks: Xi'an-Chengdu cross-city GDP, population, carbon prediction
Metrics: MAE MAPE
Baselines: Individual RP RA HSP HSA HBP HBA MMD Adv CrossTReS CARPG
Builds on: relative representations graph attention network Yabe et al. 2020 unsupervised translation
Closest papers by meaning
- Urban Region Embedding via Multi-View Contrastive Prediction (AAAI 2024 · similarity 0.86)
- LIVS: A Pluralistic Alignment Dataset for Inclusive Public Spaces (ICML 2025 · similarity 0.84)
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.84)
- Improving Region Representation Learning from Urban Imagery with Noisy Long-Caption Supervision (AAAI 2026 · similarity 0.84)
- UrbanKGent: A Unified Large Language Model Agent Framework for Urban Knowledge Graph Construction (NeurIPS 2024 · similarity 0.83)
- CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing (ICLR 2026 · 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.