Machine learning could better predict gentrification

A machine learning model developed and tested by UNSW researchers could help better give policymakers information on gentrification.

Researchers at the UNSW city futures research centre published findings noting that with better predictive data, policymakers and government can deliver more equitable city planning and outcomes.

Gentrification can displace and negatively impact residents such as low-skilled workers and vulnerable people. Governments and policymakers often struggle to address the associated harms.  

William Thackway, postgraduate researcher at UNSW City Futures Research Centre said gentrification is often identified when it's too late, and it can be costly to address the harms it has caused.

He said, “The ability of policymakers to adequately tackle harms caused by gentrification rests on proactive strategies which prevent or mitigate displacement of vulnerable people before it becomes too expensive to do so.”

Thackway, Professor Christopher Petitt, Dr Matthew Ng and associate professor Chyi Lin Lee developed the prototype machine learning model and tested various data from Sydney as a case study. 

“A key finding from our work is that the gentrification frontier is predicted to expand outwards even further from the city centre,” Thackway said.

“Previously, the rings of gentrification were in 5-10 kilometre rings around the Sydney CBD, but that is predicted to expand to 10-20 kilometres.”

The study identified an effect the researchers call ‘spill over’ as an across-the-board indicator of predicted gentrification in Sydney. ‘Spill over’ is when displaced residents from gentrification hotspots move to neighbouring suburbs where rents are slightly cheaper.

Eastwood beyond Ryde and Brookvale beyond Manly were other suburbs experiencing ‘spill over’ effects and predicted to gentrify according to this machine learning analysis.

The advantage of this new machine learning model is that it can make links between variables that are otherwise overlooked in other methods of analysis involving just human expertise, the university explained.

“Our study includes a wider range of predictor variables than previous machine learning studies, spanning socioeconomic, housing, business and Airbnb data,” Thackway said.   

The machine learning model was trained and tuned using over 80 predictor variables from a wide range of data inputs such as property reports, the census, business registry and Airbnb. 

To test its accuracy, the researchers retroactively applied the model to previously-ungentrified neighbourhoods that ended up becoming gentrified.

Family compositions and relationship status were surprisingly important indicators of gentrification in some areas of Sydney, Thackway explained.

“It was surprising to see that an increase in married couples in an area lead to a higher prediction that the area will gentrify, while areas with more divorcees and one-parent families were less likely to gentrify according to our model,” he said.

In some cases, family and relationships were as important as house prices, education and employment in predicting gentrification for a suburb.

Infancy mode

Predictive modelling and machine learning tools in the urban policy spheres are still in their infancy.

Thackway explained that there is still scepticism among policymakers about the trustworthiness of such models.

He said, “Previous machine learning models have had a ‘black box’ element to them, meaning that we can’t see how machine came to its conclusions. Because of this, the preference among policymakers is dominated by qualitative methods.” 

But this new machine learning model developed by UNSW researchers can predict gentrification with 87.3 per cent accuracy and it eliminates the ‘black box’ element by implementing a model explanation tool that interprets how the machine learning model came to its conclusions.

“Qualitative methods like the Neighbourhood Change Warning System and Gentrification Index are easy for policymakers to understand. But the downside is that they are quite simple and lack robustness,” Thackway said.

“Our machine learning model incorporates tens, if not hundreds of indicators compared to qualitative methods. The advantage of using machine learning as opposed to basic indicators in qualitative methods is the model can identify interactions and relationships between variables that one might not necessarily be able to do just from human expertise.”

The tool is in its development phase and there is scope to test it to more extreme degrees to ensure its performance. 

“Right now, the major implication of our work is that this model can produce meaningful and powerful results that will enable proactive policy decisions and interventions for urban planners,” Thackway ended.