Land-bank properties used public records
An Eye on Ohio reporting team has spent the past year working on an analysis of how cities and land banks choose to take over decrepit properties and disperse them. This effort involved hundreds public-records act requests, 5,225 lines of code and countless hours of planning, researching, programming, writing, fact-checking and editing.
Uncounted news stories chronicle the shortage of houses and rising housing prices across the country. Eye on Ohio wanted to look at the opposite end of the spectrum: What happens to the worst housing and vacant lots in Ohio cities where deteriorated houses had been demolished? How does that affect people in those neighborhoods who are struggling?
In Ohio, counties keep delinquent property lists showing which owners have not paid property taxes the previous year and how much they owe. The county auditor's website allows anyone to see property values and payment histories.
Most delinquent taxpayers eventually pay back their taxes. But Eye on Ohio started here for several reasons.
First, it would be impractical to study every property in a county to see which might be eligible for a land bank, the purpose of which is to clean up unsightly properties by getting them into the hands of owners who will maintain or redevelop them — and pay their taxes. The delinquent-taxpayer list is a public record which represents virtually all decrepit properties in a county.
Second, delinquent property owners are the biggest funders of land banks in the first place. County treasurers and prosecutors split 5% of delinquent tax revenue between them in a delinquent tax and assessment collection fund (DTAC). When a county establishes a land bank, it uses those funds. County commissioners can authorize up to 5% more.
When a taxpayer becomes delinquent, the treasurer usually sets up a payment plan. If that fails, the government sells the tax lien to a third party. But sometimes not even that is successful, particularly for abandoned properties where it can be difficult to find an owner's heirs or successors in interest.
Land banks get properties in a variety of ways. Someone can give the land bank a parcel outright or deed property in lieu of foreclosure. Typically, they remediate properties that are way behind on their taxes.
The purpose of the analysis is to determine which delinquent properties in a county go to the land bank. Each land bank has a policy that essentially says, “We try to do the best we can with our limited budget.” What exactly does that mean, mathematically?
Land banks can be effective but very limited programs. How do officials choose which of the relatively small number of properties they will foreclose upon or demolish?
During the height of the mortgage crisis, many counties got federal funds to supplement their budgets. That money is now largely gone. How will land banks treat deteriorating properties with a smaller budget?
To seek answers, Eye on Ohio looked to "machine learning," a form of artificial intelligence.
Machine learning is powerful because it flips the script on computer programming: Instead of telling the machine what's important, programmers study the data points that influence various outcomes to see what's important. Then they test for better outcomes.
Meredith Broussard notes in her book, “Artificial Unintelligence,” that “AI” is a bit of a misnomer. True artificial intelligence means computers have finally achieved consciousness. Scientists are a long way away — if that is even possible.
Why then has “artificial intelligence” become ubiquitous? Major companies and the state of Ohio refer to AI as vital to speech recognition, self-driving cars and web searches. It's essentially become shorthand for various machine-learning methods to solve a problem which a human can't easily solve.
For example: A programmer has to code image-recognition software to identify dogs in pictures. How can the programmer explain to a computer what a dog is? Chihuahuas are dogs, and so are Great Danes. But not wolves, which look a lot like dogs, or foxes.
The programmer instead could use thousands of pictures of animals handlabeled “dogs” and “foxes” and have an AI algorithm learn which are which. The computer compares patterns of each animal's eyes, nose and snout to see which sizes and shapes are a “dog.” The code tells the computer to decide a shape, such as a dog ear, is more likely a dog.
As François Chollet and J.J. Allaire wrote in their book, "Deep Learning with R," from a geometric standpoint, the computer is trying to see how to fold a piece of paper so that the maximum number of data points can be included.
Counterintuitively though, extremely high accuracy is not an end goal because of “overfitting.” A model that follows data too closely might not be good at making predictions in new data it hasn't seen before. If your dog dataset has too few Chihuahuas and not enough Great Danes, you might miss bigger dogs later.