In her work on statistical issues in criminal justice, Lum has studied uses of predictive policing—machine learning models to predict who will commit future crime or where it will occur. In her work, she has demonstrated that if the training data encodes historical patterns of racially disparate enforcement, predictions from software trained with this data will reinforce and—in some cases—amplify this bias. She also currently works on statistical issues related to criminal “risk assessment” models used to inform judicial decision-making. As part of this thread, she has developed statistical methods for removing sensitive information from training data, guaranteeing “fair” predictions with respect to sensitive variables such as race and gender. Lum is active in the fairness, accountability, and transparency (FAT) community and serves on the steering committee of FAT, a conference that brings together researchers and practitioners interested in fairness, accountability, and transparency in socio-technical systems.

“The model uses obvious predictor variables, Ball says, such as whether or not a drug lab has been busted in that county, or if the county borders the United States, or the ocean, but also includes less-obvious predictor variables such as the percentage of the county that is mountainous, the presence of highways, and the academic results of primary and secondary school students in the county.”

Kristian Lum discusses the challenges of getting accurate data from conflict zones, as well as her concerns about predictive policing if law enforcement gets it wrong.

“Every kind of classification system – human or machine – has several kinds of errors it might make,” [Patrick Ball] says. “To frame that in a machine learning context, what kind of error do we want the machine to make?” HRDAG’s work on predictive policing shows that “predictive policing” finds patterns in police records, not patterns in occurrence of crime.

As executive director of the Human Rights Data Analysis Group, Megan Price uses statistics to shine the light on human rights abuses.

Kristian Lum, lead statistician at the Human Rights Data Analysis Group (and letter signatory), fears that “in order to flag even a small proportion of future terrorists, this tool will likely flag a huge number of people who would never go on to be terrorists,” and that “these ‘false positives’ will be real people who would never have gone on to commit criminal acts but will suffer the consequences of being flagged just the same.”

One of the researchers, a Michigan State PhD candidate named William Isaac, had not previously heard of New Orleans’ partnership with Palantir, but he recognized the data-mapping model at the heart of the program. “I think the data they’re using, there are serious questions about its predictive power. We’ve seen very little about its ability to forecast violent crime,” Isaac said.

Kristian Lum is focusing on artificial intelligence and the controversial use of predictive policing and sentencing programs.

What’s the relationship between statistics and AI and machine learning?

AI seems to be a sort of catchall for predictive modeling and computer modeling. There was this great tweet that said something like, “It’s AI when you’re trying to raise money, ML when you’re trying to hire developers, and statistics when you’re actually doing it.” I thought that was pretty accurate.

On the heels of the Women in Data Science conference, HRDAG executive director Megan Price says, “I think creativity and communication are probably the two most important skills for a data scientist to have these days.”

Researchers from Rice University and Duke University are using the tools of statistics and data science in collaboration with Human Rights Data Analysis Group (HRDAG) to accurately and efficiently estimate the number of identified victims killed in the Syrian civil war.

…

Using records from four databases of people killed in the Syrian war, Chen, Duke statistician and machine learning expert Rebecca Steorts and Rice computer scientist Anshumali Shrivastava estimated there were 191,874 unique individuals documented from March 2011 to April 2014. That’s very close to the estimate of 191,369 compiled in 2014 by HRDAG, a nonprofit that helps build scientifically defensible, evidence-based arguments of human rights violations.

But while HRDAG’s estimate relied on the painstaking efforts of human workers to carefully weed out potential duplicate records, hashing with statistical estimation proved to be faster, easier and less expensive. The researchers said hashing also had the important advantage of a sharp confidence interval: The range of error is plus or minus 1,772, or less than 1 percent of the total number of victims.

“The big win from this method is that we can quickly calculate the probable number of unique elements in a dataset with many duplicates,” said Patrick Ball, HRDAG’s director of research. “We can do a lot with this estimate.”