Abstract
The H4M dataset provides a massive, multi-modal window into the socioeconomic pulse of Beijing by integrating real estate transactions, points of interest (POI), traffic patterns, and over 100 million geolocated microblog posts. This paper describes the dataset's composition and its potential as a comprehensive "test bed" for researchers to model urban dynamics and socioeconomic indicators with unprecedented granularity.
Introduction
Understanding the socioeconomic health of a city traditionally relied on static surveys or single-source datasets, such as traffic logs or housing prices. However, modern urban environments are complex systems where social sentiment, mobility, and commerce are deeply intertwined. Existing datasets in this field often suffer from being outdated, limited in scale, or lacking geographic coordinates. The H4M dataset bridges this gap, providing a "multi-view" perspective that captures how Beijing moves, spends, and feels.
The H4M Dataset
The H4M dataset is composed of four distinct pillars, each offering a unique lens on the city:
- Real Estate: 28,645 entries (2017–2018) including 21 variables such as price, building year, and internal layout.
- Points of Interest (POI): 497,256 entries across 18 categories, including restaurants, hospitals, and transit hubs.
- Traffic Flow: Data for 23,772 active grids recorded at 5-minute intervals.
- Microblogging: Over 100 million geolocated Sina Weibo posts (2013–2015).
Methodology
The researchers employed extensive web crawling and data cleaning to build H4M from four primary sources: Lianjia for real estate, Baidu Maps for POIs and traffic, and Sina Weibo for social media. A key feature of this work is the "multi-distributional" approach. Rather than forcing all data into a single format, the authors maintain the intrinsic characteristics of each source—using JSON for structured property data, Pickle files for high-frequency traffic grids, and plain text for social media posts.
Spatial Analysis and Insights
A primary contribution of H4M is the revelation of spatial heterogeneity across data types. The distributions of these four pillars are not uniform; real estate and POIs cluster differently than traffic or social media activity. This suggests that a model relying on only one source would provide a biased view of the city's structure.
Validation through visualization confirms the dataset's real-world accuracy. For instance, real estate prices are significantly higher in the city center and lower in the suburbs, reflecting expected economic pressures. Similarly, traffic data captures the expected diurnal cycle—high speeds in the early morning and significant drops during the day.
Conclusion and Future Work
While H4M is a powerful tool, it faces a limitation in temporal misalignment: microblog data spans 2013–2015, while real estate and traffic data are from 2017–2018. Researchers must account for urban changes occurring in that gap. Future work should focus on multi-modal fusion, using representation mapping to predict large-scale traffic congestion or refine property valuation models by incorporating the "emotional" and "functional" data provided by social media and POIs.
Summary updated