In this short talk, I am going to speculate on how we might begin to explain urban development patterns in terms of city shape, density and location as a function of a series of independent factors that when combined yield the kinds of patterns that we observe in cities of different sizes and forms. In short, I will explain how we can derive a standard set of factors and their relationships to one another in a such a way that when we observe them for any town, we are able to use these relationships to explain the morphology of the place in terms of shape, density and location. The relationships are generated from the neural nets that link the factors to the town’s morphology. In terms of segregated urban patterns, it may be possible to do the same to explain how different patterns of segregation in terms of location emerge as a consequence of the relationships amongst a relevant set of factors but I will only go as far as speculating that this approach appears promising.

Now to introduce the idea, I am going to turn the problem on its head. In the design science literature which is still rather rudimentary, there are models of how factors influencing locational designs can be applied in analogy to the way we reduce conflict between these factors thus yielding compromise solutions. These date back many years to work by Christopher Alexander, Herbert Simon and so on but have remained very much underground, notwithstanding their links to the form of strong AI that began the field. If you think of the problem network as one in which a series of individuals hold views about the best location for urban development, then the network determines who interacts with who. When a process of rational compromise is associated with the transmission of the set of best locations between the actors, then in many circumstances the ultimate compromise solutions where all the actors agree can be generated as the steady state of this process.

This process is akin to generating a network which implies a weighting of individual responses based on the different factors that leads to a stable pattern. It is but a short hop to the idea that if we have many urban development patterns – many solutions – and many factors that we can associated with these solutions, we can engage in a process of iterative weighting that ultimately produces a stable set of weights that for any set of factors will produce a pattern of urban development that it close to those we observe. This in essence is a stable neural net that can then be used for situations where we do not know the urban development pattern but we do have a series of independent factors. This clearly might relate to a predictive context.

In this very short talk, I will outline the problem, and then introduce the various network methods used for reconciling conflicts in design – that is generating design solution in terms of urban development – and then speculate on generalising the method by throwing at the net, many different factors associated with many different patterns. Hopefully this will produce a weighted net that explains many such patterns to a degree good enough to know that the set of factors we have produced and their weights explain enough of urban development for the system to be useful. It is not unlike any machine learning problem based on neural nets that involves many determinants used to recognise the different ‘faces’ of urban development. What is nice about this interpretation is that it opens the way up to figuring out what the neural net actually means in terms of the way individuals might relate the set of factors to explaining actual urban development: to the social process of design and development.

References

M. Batty (2013) A New Science of Cities (Part 3), The MIT Press, Cambridge MA; M. Batty (2020) Complexity in Design: Optimal Location Through Spatial Averaging, in G. de Roo, C. Yamu, and C. Zuidema (Editors) Handbook on Planning and Complexity, Edward Elgar, Cheltenham, Glos UK, 302-317; at http://spatialcomplexity.blogweb.casa.ucl.ac.uk/files/2020/08/Complexity-in-Design-Batty-2020.pdf

As we live in an era of smart cities, big data and algorithms there are increasing a number data- driven tools to support planner and policy-makers in their quest shape more liveable cities. A number of these digital planning tools are based on big data, machine learning and artificial intelligence methods and techniques. Yet one of the key challenges is to incorporate such data and analytics into digital planning tools used in practice to ultimately make better planning decisions. The implementation gap of tools in practice is one where many researchers are focusing their efforts (Brömmelstroet & Schrijnen 2010; Geertman 2017; Pettit et al. 2018). These studies are encapsulated more broadly in the planning support system (PSS) literature (Geertman et al. 2019). The advantage of many PSS tools is that they have been underpinned by simple or transparent modelling paradigms such as the regression approaches used in hedonic price models (HPM). Such approaches enable the coefficients of the models to be exposed and further analysed. For example, the distance to train stations is a common variable used in a number of HPM studies, from which an economic value can be determine in how much this contributes to a property value (Du & Mulley 2006, Lieske et al. 2019). With the advent of big data such databases supporting HPM studies can now be developed from millions (or billions) of data points. Likewise, with the temporal nature of such database of property prices going back well beyond 10 years, they provide fertile ground for running Machine Learning and Artificial Intelligence algorithms. Given this context, this presentation will focus on efforts underway in developing a suite of automated valuation models (AVMs) being applied and tested in Sydney and across Australia. In this presentation I will outline the models, and the challenges and opportunities which arise in the use of different automated valuation models for predicting residential property prices. Also, I will reflect on how these models can underpin the next the generation of digital planning tools.

References

Brömmelstroet, M. T., & Schrijnen, P. M. (2010). From planning support systems to mediated planning support: a structured dialogue to overcome the implementation gap. Environment and Planning B: Planning and Design, 37(1), 3-20.

Du, H., & Mulley, C. (2006). Relationship between transport accessibility and land value: Local model approach with geographically weighted regression. Transportation Research Record, 1977(1), 197-205.

Geertman, S., Zhan, Q, Allan, A., Pettit, C.J (2019) Computational Urban Planning and Management for Smart Cities Springer International Publishing.

Geertman, S. (2017). PSS: Beyond the implementation gap. Transportation Research Part A: Policy and Practice, 104, 70-76.

Lieske, S. N., van den Nouwelant, R., Han, J. H., & Pettit, C. (2019). A novel hedonic price modelling approach for estimating the impact of transportation infrastructure on property prices. Urban Studies, 0042098019879382.

Pettit, C., Bakelmun, A., Lieske, S. N., Glackin, S., Thomson, G., Shearer, H., ... & Newman, P. (2018). Planning support systems for smart cities. City, culture and society, 12, 13-24.

Prediction of displacement becomes a challenging task where displacement has already occurred in the past and exclusion has become entrenched. For example, no loss of low income earners, or gains of high income earners can be measured, since a neighbourhood could be already populated with high income earners and hence present a high entry barrier to low income earners. In such a case, what could be more reliably measured is whether the concentrations of low or high income neighbourhoods has been increasing over time. We use Shannon’s information theory based entropy and spatial entropy measures (Batty, 1974) to examine whether the concentrations of neighbourhoods by income has increased, using Sydney as our demonstration case. This work extends the neighbourhood porosity measure that was presented in the first symposium in this conference series at Berkeley.

Researchers have long struggled in using secondary census data to measure neighborhood change, specifically in the form of gentrification and displacement. Though some researchers have devised typologies of neighborhood change that predict future transformation, for instance using machine learning to assign gentrification risk to neighborhoods, their predictive power remains questionable, perhaps in part because of the use of census data that is out-of-date or unreliable at a fine geographic scale (Chapple and Zuk, 2016; Reades, De Souza and Hubbard, 2018). Moreover, most analyses focus on gentrification rather than displacement, due to the greater ease of measuring the influx of capital and/or high-educated newcomers than forced moves or exclusion.

Might real-time data on activity patterns improve the accuracy of these models by pinpointing the areas of dynamic change? In this paper I develop typologies of neighborhood change in the 2000s for four regions: the San Francisco Bay Area, New York, Buenos Aires, and Bogota. After identifying areas which have not gentrified, but are considered at risk (based on loss of low-income households and increase in property values), I validate and refine the models using geotagged tweets from 2012 to 2015 (based on the methodology in Shelton, Poorthuis, and Zook 2015). I expect to find that real-time geographic information refines and narrows the set of neighborhoods considered at risk for gentrification and displacement via conventional data. I conclude by exploring ways to incorporate these data into our understanding of neighborhood change – and policy-making for cities -- on a more real-time basis.

References

Chapple, Karen and Miriam Zuk. 2016. “Forewarned: The Use of Neighborhood Early Warning Systems for Gentrification and Displacement.” Cityscape 18,3: 109-130.

Reades, J., De Souza, J., & Hubbard, P. (2018). Understanding urban gentrification through machine learning. Urban Studies, 0042098018789054.

Shelton, Taylor, Ate Poorthuis, and Matthew Zook. "Social media and the city: Rethinking urban socio-spatial inequality using user-generated geographic information." Landscape and Urban Planning 142 (2015): 198-211.

Gentrification has been a key issue for governments over the past few decades with it affecting cities around the world but measuring, and in particular, predicting it is difficult. Data from governments on social demographics in cities have been found to be important to measuring gentrification, but these datasets normally arrive after a long time lag. The availability of new data from digital platforms sources provides an opportunity to revisit our understanding of changing neighbourhoods and enable new measures of these changes in close to real-time. In this paper, we ask this question: how might activity patterns on the digital platform Hemnet improve the understanding of two types of neighbourhood change – gentrification and exclusion – by identifying dynamic areas across scales for the city of Stockholm? These results could provide an early warning system to urban decision-makers and communities to make changes to improve the social dynamics in Stockholm and to prevent the displacement of current residents.

First, we develop a typology of neighbourhood change from 2013 to 2019 for Stockholm by using time series clustering to find if a neighbourhood is ascending, declining, or is stable. From the ascending neighbourhoods we then undertake a second time series clustering analysis, this time to reveal different types of gentrification: super gentrification, marginal gentrification and classical gentrification. Next, we collect data from Hemnet, which is updated daily with real-estate price sales in Stockholm. The platform has data ranging from 2013-2019 which we use in our measures. With this data, we built a time series deep neural network model to predict gentrifying neighbourhoods in Stockholm for 2024. We do a second prediction this time focusing on the different typologies of gentrification. As a consequence, we are able to uncover the most significant factors that predict an early warning system for neighbourhood change.

The aim of this research is to reveal the potential to analyse neighbourhoods at a fine-grain level to understand micro-macro complex social neighbourhood relationships. Creating an early warning system for Stockholm in real-time can create more socially inclusive neighbourhoods and prevent residents from being displaced due to gentrifying and excluding areas. Our results can help provide a tool for Stockholm officials to implement policies geared towards the needs of community residents; these policies can range from reserving affordable housing units to educating residents of their renting rights, to helping small businesses negotiate long-term lease extensions. ​

Keywords: data-driven urban studies, gentrification, neighbourhood trajectories, deep learning, Stockholm, principal component analysis, Hemnet, LSTM, time-series clustering