Researchers propose a graph-based machine learning method to quantify the spatial homogeneity of subarrays

This Article Is Based On The Research Paper 'Quantifying spatial homogeneity of urban road networks via graph neural networks'. All Credit For This Research Goes To The Researchers 👏👏👏

Please Don't Forget To Join Our ML Subreddit

Researchers from Purdue University and Peking University recently completed a study using machine learning methods to better understand road networks in cities around the world. Their research, which was published in Nature Machine Intelligence, details the results of a survey based on data from road map data collected from 30 cities around the world.

According to the researchers, urban road networks (URNs) are the economic engine of a city. They are formed by many socio-economic elements (including population, economy) and the history of urban development. They are essential for human mobility, healthy movements, the spread of biological viruses and the creation of pollution. Traditional road network measurements based on simple parameters, on the other hand, provide only an approximate description of URNs.

Previous research has suggested that the spatial homogeneity of URNs follows a pattern. Graph-based neural networks (GNNs) are advanced graph-based machine learning methods commonly used in computer vision and natural language processing applications that could capture these patterns. GNNs can learn network representations from very large volumes of data. The research team used GNNs to assess more than 11,790 URNs in 30 cities around the world to predict a new parameter called network homogeneity in their research.

The researchers explained that many planners have looked at cities using case-by-case methodologies. They had sought to use the power of machine learning technology and insights from big data to gain a global understanding of the urban system, which included cities in both developed and developing countries. However, using global data, quantitative comparisons of the urban environment between cities are limited.

The researchers separated all of the URNs in their dataset into two parts before completing their analyses: the “hidden region” and the “observed region.” Next, they trained the GNNs to learn the patterns of road network structure in the observed regions, allowing them to predict the network structure in the hidden region.


The researchers defined the metric they studied, “network homogeneity,” as the F1 score of model performance in predicting the hidden region in URN data in their paper. F1 scores represent the precision and recall accuracy of classifiers or machine learning systems. A higher F1 score meant that the model was more likely to successfully infer the hidden region from the region seen in the context of the team’s search. It also meant that the URNs were more homogeneous.

Top-down planning policies often govern the evolution of road networks in major cities. These policies could be evaluated and compared using the researcher’s graphical neural network-based model and metric.

The researchers’ analysis revealed links between URNs, a country’s gross domestic product, and population growth and offered a statistic that could aid in the evaluation of urban planning initiatives. The results confirm the intimate relationship between human activities and urban environments.

The research team’s findings helped them understand the complex interplay between many components of the urban system. Their study is unusual in that it combines machine learning with urban science: the F1 score is a typical machine learning metric, while homogeneity characterizes the network structure of URNs.

According to the research group, they were the first to examine whether machine learning algorithms could meaningfully study URN systems. Their findings show that advanced machine learning algorithms could be used to collect rich data on socio-economic aspects and the evolution of the city over time.

Imagine walking around a corner and predicting what the next block will look like just by looking at the neighborhood you just passed through. When you travel to a new city, you may feel very familiar with the surroundings and believe it is the same as another place you know. This can happen in different cities of the same country or in several countries. Interestingly, this occurrence is not random but rather can be explained by simple things like intra-urban and inter-urban homogeneity and can be traced back to the urban planning culture of Europe, South America North and Asia, from ancient cities to modern cities.

The neural graph networks built by this group of scholars could eventually be used in various countries around the world to compare cities, evaluate policies and summarize activities. Surprisingly, the model presented in the latest research could be extended to study larger urban regions and examine changes over longer periods of time.

The researchers shared that further studies may look at the road networks of cities around the world and of different sizes, while their analysis looks at 30 major cities. In addition, they can assess and evaluate the homogeneity of a road network over time. Homogeneity theories of street views, land use, and other infrastructure networks could be constructed in addition to road networks. There is a much greater chance of knowing more about complex urban processes.

The team’s new study is one of the first to use advanced graphical neural networks to study URNs. The researchers hope to improve their model and apply it to more data in future studies to learn more about URN homogeneity. They are currently conducting research on URNs and their links to various socio-economic characteristics.

The researchers want to use machine learning models to look at other forms of data collected in urban contexts and look at how URNs change over time. For example, they want to simultaneously examine street photographs, mobility interactions, and internet browsing data to find more complex patterns that affect people’s lives in cities. For example, it could help to better understand social inequalities and regional poverty.

Finally, the researchers want to undertake more research on the overall potential of AI in urban science. Their study could motivate other research organizations to use machine learning in urban science, leading to new insights into the history and evolution of cities around the world.



Sherry J. Basler