City2City: Translating Place Representations across Cities

Takahiro Yabe, Kota Tsubouchi, Toru Shimizu, Yoshihide Sekimoto, Satish V. Ukkusuri

Abstract

Large mobility datasets collected from various sources have allowed us to observe, analyze, predict and solve a wide range of important urban challenges. In particular, studies have generated place representations (or embeddings) from mobility patterns in a similar manner to word embeddings to better understand the functionality of different places within a city. However, studies have been limited to generating such representations of cities in an individual manner and has lacked an inter-city perspective, which has made it difficult to transfer the insights gained from the place representations across different cities. In this study, we attempt to bridge this research gap by treating cities and languages analogously. We apply methods developed for unsupervised machine language translation tasks to translate place representations across different cities. Real world mobility data collected from mobile phone users in 2 cities in Japan are used to test our place representation translation methods. Translated place representations are validated using landuse data, and results show that our methods were able to accurately translate place representations from one city to another.

Introduction

In recent years, urban planners and data scientists have utilized "place embeddings"—mathematical vectors that represent the functional characteristics of a neighborhood based on human mobility patterns. Much like word2vec identifies semantic relationships between words, these embeddings reveal how specific grid cells function as shopping hubs, residential zones, or business districts.

However, a significant limitation exists: these embeddings are city-specific. Because the vector space for one city is generated independently of another, they cannot be directly compared. This "silo" effect prevents the transfer of critical insights—such as disaster evacuation strategies—between cities. The City2City framework addresses this by treating cities like languages and aligning their disparate vector spaces through unsupervised translation techniques.

The City2City Framework

The authors propose a three-step pipeline to bridge the gap between independent city representations:

  1. MobLSTM: A Long Short-Term Memory (LSTM) network is trained to predict a user's next "staypoint" based on mobility history. The weights of the embedding layer serve as the initial place representations.
  2. Translation Approaches: Three methods were tested to align the embeddings of two different cities ($\phi$ and $\psi$):
  3. Joint-MobLSTM: Training a single model on merged data from both cities.
  4. MobLSTM-P (Procrustes): Learning embeddings separately and finding an orthogonal rotation matrix to align them using popular locations as "anchor points."
    • MobLSTM-Adv (Adversarial): A GAN-based approach where a discriminator attempts to identify the source city of a vector while a mapping function tries to align them.
  5. Validation: The translated points are compared against 17-category land-use data to verify if they align with real-world functions.
The City2City framework translates place representations from different cities into a common vector space to enable cross-city functional analysis.

Experimental Setup and Datasets

The study utilized high-resolution datasets from two Japanese cities:

  • Kumamoto: Approximately 94,000 users.
  • Okayama: Approximately 119,000 users.

Mobility data was provided by Yahoo Japan (GPS staypoints), while ground-truth functional labels were derived from the Japanese government's "Urban Area Land Use Mesh Data," covering 100m x 100m grid cells.

Key Results and Insights

The core finding of the study is that cities can indeed be treated like languages. If a "trip" is a "sentence" and a "place" is a "word," then moving between cities becomes a translation task.

Quantitative results, measured by Average Mutual Cosine Similarity (AMCS) and Mutual Norm Distance, demonstrated that MobLSTM-P (Procrustes) significantly outperformed other methods. It successfully mapped shopping districts and residential areas in Okayama to their functional counterparts in Kumamoto. Interestingly, the Joint-MobLSTM approach performed poorly; the model became too proficient at distinguishing between the two cities, leading it to embed them in entirely separate regions of the vector space rather than finding commonalities.

Quantitative comparison of translation methods across different land-use labels. MobLSTM-P (maroon) consistently shows the highest mutual cosine similarity and lowest norm distance, indicating superior alignment.

Popularity proved to be a vital anchor. The most frequently visited locations (train stations, major malls) served as universal reference points for alignment. However, the model struggled with farmlands due to the lack of distinct "staypoint" patterns in those areas, resulting in noisy embeddings.

Conclusion

The City2City framework proves that unsupervised translation can effectively align functional urban representations across different cities. While this study focused on culturally similar Japanese cities, future work will explore translating between vastly different urban structures (e.g., Tokyo vs. Western cities) and incorporating additional data sources like satellite imagery. The ultimate goal is to enable the transfer of predictive models for urban optimization and disaster response to cities where local data may be scarce.

Summary updated

Abstract

Large mobility datasets collected from various sources have allowed us to observe, analyze, predict and solve a wide range of important urban challenges. In particular, studies have generated place representations (or embeddings) from mobility patterns in a similar manner to word embeddings to better understand the functionality of different places within a city. However, studies have been limited to generating such representations of cities in an individual manner and has lacked an inter-city perspective, which has made it difficult to transfer the insights gained from the place representations across different cities. In this study, we attempt to bridge this research gap by treating \textit{cities} and \textit{languages} analogously. We apply methods developed for unsupervised machine language translation tasks to translate place representations across different cities. Real world mobility data collected from mobile phone users in 2 cities in Japan are used to test our place representation translation methods. Translated place representations are validated using landuse data, and results show that our methods were able to accurately translate place representations from one city to another.
BibTeX
@inproceedings{yabe2019city,
  author = {Takahiro Yabe and Kota Tsubouchi and Toru Shimizu and Yoshihide Sekimoto and Satish V. Ukkusuri},
  title = {{City2City: Translating Place Representations across Cities}},
  booktitle = {SIGSPATIAL/GIS},
  year = {2019},
  url = {https://openreview.net/forum?id=rzzlCKYHo5},
}
Figure