Function2vec: A Geographic Knowledge Graph Model of Urban Function Evolution and Its Application

Tianle Li, Minrui Zheng, Xiaoli Wang, Xinqi ZhengISPRS Int. J. Geo Inf 2023 · DOI

TL;DR

This paper introduces Function2vec, a novel model that uses urban tree theory and graph embeddings to visualize and analyze the spatial correlations of urban function evolution (UFE). By representing UFE as a network, Function2vec offers a more nuanced understanding of how cities change, moving beyond simple density maps to capture complex interdependencies.

What Problem Does This Paper Address?

Understanding how urban functions change over time (Urban Function Evolution - UFE) is crucial for sustainable urban planning, especially in rapidly developing "emerging cities." Existing methods often focus on land expansion or analyze functional density, but they struggle to visually represent the spatial correlations between different evolving functions. This makes it difficult to grasp how one function's change might influence or be influenced by another in a spatially interconnected way. The paper aims to bridge this gap by providing a theoretical framework (urban tree theory) and a computational model (Function2vec) to capture these spatial relationships.

Methods

The core of this work is the Function2vec model, which combines urban tree theory with a modified node2vec graph embedding technique.

  • Urban Tree Theory: This theory views urban systems as having a hierarchical, tree-like structure. In this context, the "root" represents an early-phase urban function, and the "fruit" represents a later-phase function. The model categorizes UFE into three types:
  • Function Replace (FR): An early function is replaced by a later one.
  • Function Newly Added (FNA): A new function appears.
  • Function Vanishing (FV): An existing function disappears.
  • Function2vec Model:
    1. UFEN Construction: The model first constructs an Urban Function Evolution Network (UFEN) based on the urban tree theory. This network represents the relationships between different UFE patterns.
  • Node2vec Modification: The standard node2vec algorithm, which learns embeddings for nodes in a graph, is adapted. Here, the "nodes" are the urban functions themselves. The algorithm uses sampling strategies (Depth-First Search and Breadth-First Search) to generate sequences of nodes representing functional evolution paths.
  • Embedding and Correlation: These sequences are then fed into a skip-gram model to generate feature vectors (embeddings) for each functional node. The Pearson correlation coefficient is used to calculate the similarity (correlation) between these embeddings, quantifying the spatial correlation between UFE patterns.

Datasets

The study utilizes Point-of-Interest (POI) data for Shenzhen city, China, from two time points: 2013 and 2022. This data was obtained from the Amap platform. The POI data was classified into urban functions based on categories. No new benchmark datasets were introduced; rather, existing POI data was leveraged.

Key Insights and Ideas

  • Visualizing Spatial Correlation: The most significant contribution is the ability to visually express the spatial correlation of UFE patterns using a tree-like structure. This moves beyond simple density maps to show how functions evolve together or in relation to each other across space.
  • Urban Tree Theory as a Framework: Applying urban tree theory provides a natural and intuitive way to conceptualize the hierarchical and interconnected nature of urban function changes.
  • Function2vec as a Powerful Tool: The Function2vec model effectively integrates graph embedding techniques with urban theory to quantify and visualize these complex relationships, offering a novel approach to urban analysis.
  • The 20-80 Law in UFE: The study found that Function Vanishing (FV) and Function Newly Added (FNA) patterns exhibit a long-tailed distribution, aligning with the 20-80 law, suggesting that a small number of functions account for a large proportion of these changes.

Key Results

  • Dominant UFE Pattern: From 2013 to 2022, Function Replace (FR) was the dominant UFE pattern in Shenzhen, accounting for 89.44% of changes.
  • "Multi to Multi" Dominance: Within FR, the "multi to multi" sub-pattern (where mixed functions evolve into other mixed functions) was the most frequent, indicating complex interdependencies.
  • Long-Tailed Distribution for FNA/FV: Approximately 80% of functions fall into the "tail" of the distribution for FNA and FV, meaning a large proportion of functions have low link values in these categories.
  • Healthcare Services: Healthcare services are well-suited for complementary relationships with other functions, reinforcing their importance.
  • Science, Education, and Culture: These functions require a higher degree of complementarity with other urban functions.
  • Administrative Offices: These exhibit diverse evolutionary patterns.
  • Transportation Hubs: When integrated with other functions, transportation hubs lead to significantly deviating urban function evolution from their original patterns, acting as auxiliary functions that influence growth.

Limitations and Future Work

  • Data Limitations: The study primarily relies on POI data. The authors suggest that incorporating multi-source data could enrich the expression of UFE.
  • Single City Case Study: The model was applied to Shenzhen. Future work could involve comparing and analyzing multiple cities to validate and generalize the findings.
  • Urban Network Theories: Urban tree theory is one of many urban network theories. The authors propose exploring other theories to further enhance UFE expression.
  • Model Integration: The authors suggest integrating this model with other urban data samples and using the results to simulate UFE in later stages.

Summary updated

BibTeX
@article{li2023function,
  author = {Tianle Li and Minrui Zheng and Xiaoli Wang and Xinqi Zheng},
  title = {{Function2vec: A Geographic Knowledge Graph Model of Urban Function Evolution and Its Application}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2023},
  doi = {10.3390/ijgi12110458},
}

Loading related content...

Figure