HouseUP identifying a property that is already slated for development, a good sign in terms of the statistical model accuracy.
Input data on demographics for Atlanta (left) and the deep learning models prediction for where affordable housing should be (right).
Atlanta Mayor Keisha Lance Bottoms announced in April a $60 million increase in budget to be used towards developing new affordable housing. Finding appropriate locations to build new affordable and low-income housing is a more challenging problem that initially meets the eye. According to Vox 2016, “People think [low-income housing] will further exacerbate the difficulties of neighborhoods already saddled with difficulties, but the data shows otherwise. Crime rates fall and property values rise when subsidized housing is built in a poor neighborhood.” Low-income housing has been seen to have a positive impact on the surrounding community if implemented in the right location. A 2016 study from Rebecca Diamond and Timothy McQuade of Stanford Business School had similar findings: "Lowering crime in low income areas appears to be one of the driving mechanisms through which LIHTC improves low income neighborhoods.”
In addition, simply picking the location that will result in the lowest rent is rarely the best option. Living outside of a metropolitan area is often far cheaper, but the cost of transportation into the city can be extremely prohibitive. As Vox puts it living far from the city "increases their transportation costs — so a simple ‘affordable housing’ metric might not capture the whole story”. To combat this issue, The Center for Neighborhood Technology developed the H+T Affordability Index, which factors both housing and transit costs into the cost of living in an area (something which is often ignored during new low-income housing building).
We made HouseUP to help city planners and government officials make informed decisions about where to place affordable housing.
How HouseUP helps
HouseUP aggregates geographic data based on US Census polling, the H+T Affordability Index, and a few other sources. Using our own statistical models and analysis software provided by Esri's ArcGIS suite, HouseUP reports the areas that are optimal for building low-income housing.
Accomplishments that we're proud of
HouseUP's statistical model picked a few spots that already have successful low-income housing projects. We think this is a great indicator that it is correctly predicting other locations that housing could be built. It is successfully taking the guess work out of location scouting.
Additionally, we developed heatmaps of Atlanta that we used to train models of the pix2pix neural net. The ground truth heatmap was created by processing 17,000 data points into a square heatmap image representing the number of low-income housing units built at given locations from 1992 - 2015. The predictive data was processed by creating another square heatmap image with 17,000 data points representing median household income, unemployment, and current units of low-income housing. These two squares were then each sliced into 400 segments that we fed into the neural net to train the models, which were used to predict future project locations.
How we built HouseUP
3 Cups React. 4 Pounds of ArcGIS. 1 Gallon of Coffee. A dash of machine learning. A pinch of Python data processing.
Stir vigorously in a large mixing bowl.
Challenges we ran into
Turns out US Census Data is VERY large and hard to work with. Getting data into the appropriate formats was more time consuming than we anticipated. Because of this we were only able to try one iteration of data processing and training of our model.
What we learned
Esri ArcGIS! Very powerful.
What's next for HouseUP
We would like to add a civilian facing application that improves civic engagement. We think people would be more willing to help fund low-income housing projects if they are local to them. We would like to offer a way for citizens to donate to specific construction projects in their area that they want to support. Not only does this help with funding, but also with getting the cities population involved in the development of low-income housing.
For the Machine learning portion of our project we were sadly only able to do one iteration of our data and training. We did, however, learn a lot about how we can make this model better in the future. For one we need to limit both our Ground Truth data and our input data to more recent years, as well as limit it to a much closer map of Atlanta (or whatever city we are studying). Another improvement would be to have more data in our input.