New Open Source Machine Learning Tools Help Prevent Child Abuse
A pro-active approach
Limited funding for child welfare services and related prevention efforts forces difficult choices and can have tragic consequences. It has also contributed to a climate in which policy decisions regarding the safety of children and families are often reactive—made in response to incidents or particular cases—and may respond to a child’s need too late. A newly released open source predictive tool from Urban Spatial and Predict-Align-Prevent (PAP) seeks to provide a more proactive and innovative approach.
The tool aims to keep more children and families safe by providing appropriate stakeholders with predictions about areas where maltreatment is concentrated, which can then be used to better align community resources with the demand. Ken Steif, founder of Urban Spatial and Director of the Master of Urban Spatial Analytics program at Penn’s Weitzman School of Design, calls it “a framework that embeds machine learning into the strategic planning process – in this case, for child welfare.”
This geospatial approach is also a key tenet of the three-part continuous quality improvement cycle employed by PAP nationally to uncover, strengthen, and replicate effective prevention efforts in order to address and alleviate child maltreatment.
A place-based approach
PAP and Urban Spatial’s approach to predicting risk in child welfare is place- rather than people-based, which makes this tool novel in the field. In a people-based approach, individual- and household-level data would be linked across domains (i.e., education, criminal justice, housing, etc.) to generate a risk score interpreted as predicting the probability that abuse is happening in the present moment; resources are then allocated at the household- and individual- level. In contrast, place-based approaches gather de-identified, geospatial data on abuse events and environmental characteristics (i.e., where are children experiencing the greatest exposure to a series of risk and protective factors) to estimate a score interpreted as the geospatial risk for maltreatment in an ~1000 square foot area, and resources are then allocated at the community level.
The place-based approach also differs in its response to typical privacy concerns about data use. While the methodology of spatial analysis employed by the predictive tool may mirror methods used for “predictive policing,” PAP & Urban Spatial created custom bias metrics that are embedded directly into the framework. “We’ve done a good job in opening up the black box to make it possible to see when and whether some of these biases exist,” Steif said, emphasizing the importance of making predictive tools open source. Backing this up, a recent ethical analysis of the tool found that “on balance, the benefits of the PAP program would outweigh the risks it poses.” The external ethics review can be read in full here.
The first iteration
Working with the Virginia Department of Social Services in the City of Richmond—a jurisdiction that saw over 6,400 cases of maltreatment between July 2013 and 2017—PAP and Urban Spatial first deployed their proactive, place-based approach and geospatial machine learning tool with the goal of preventing child maltreatment.
Dyann Daley, founder and CEO of PAP, speaking at a recent Community Kickoff online conference to promote the use of the learning tools, shared three questions that were central to the efforts in Virginia:
How can we find at-risk children before they are injured?
What do we do once we find them?
And, are the aligned services, which are designed to support children who are at-risk or experiencing maltreatment, making a difference?
What they found validated both the PAP approach, and the value of the new machine learning tool.
In certain areas, concentrated exposure to adverse experiences increased the spatial risk of maltreatment, with domestic assaults being the primary experience that increased risk in the area studied.
More generally, Steif explained that, at this stage, the tool “borrows the observed maltreatment experience and test[s] how generalizable that experience is to other places where maltreatment has yet to be reported.”
Next, the team evaluated the spatial distribution of existing resources. They employed a gap analysis to identify places with high risk and low resources, and found that a majority of resources were concentrated in downtown Richmond, despite greater risk in outlying neighborhoods (pictured above).
With these ratios and visual tools in hand, stakeholders in Virginia were primed to advocate for changes in resource distribution in order to reach more children in need, and leverage community partnerships to keep more children safe.
A more thorough walkthrough of the tool’s development and methodological basis is available through the recordings of online conferences and related slides, which outline the following:
Next steps for scale
Building on the success of their work in Richmond, Urban Spatial and PAP are now deploying this solution to other jurisdictions. They plan to continue to refine the codebase as an R package to make it more easily accessible, while also offering additional education materials, a book and curriculum to help jurisdictions develop skills and capacity. Exploratory analysis can be conducted with only basic GIS skills, helping non-technical decision makers visualize and understand the relationship between child maltreatment and measures of exposure. An entry-level GIS analyst using these tools also has the power to take predictions, create a map, and output a spreadsheet demonstrating the locations throughout the city best suited for education and prevention programs.