Machine learning takes on college dropouts

If AI-driven machines can defeat the world’s greatest chess players and, even more improbable, the globe’s premier Go strategist, what chance does a college dropout have against machine learning technology? Slim to none, predicts one university research director.

Sudha Ram, a Professor of Management Information Systems and Director of the Center for Business Intelligence and Analytics with the University of Arizona, is leading a research project at UA that aims to prevent college dropouts from dropping out in the first place.

Ram’s efforts are nothing new for U.S. colleges and universities. Many schools use predictive analytics to help reduce freshman attrition rates. UA, for example, already tracks some 800 data points toward this effort. What makes Ram’s research unique are the types of data being collected and how those metrics are analyzed to more effectively identify at risk students.

The first several months of freshman year are the most harrowing for students. Colleges and universities know this. They also know that there are a number of early indicators for students who are most at risk for leaving after their first year. Most obvious are first-semester grades, financial aid activity, and students’ participation in course management systems. But even that information may come too late to make a difference. (Research suggests that most freshman make the decision to leave school within the first 12 weeks.)

Less evident but infinitely more powerful, says Ram, are social- and behavioral-related metrics such as shrinking social networks, fewer social interactions, and less-established routines.

Ram’s stockpile of student activity data comes from the university’s ID card tracking system, which collects information on everything from what students buy and eat to the buildings and spaces they frequent.  Using large-scale network analysis and machine learning techniques to crunch three years worth of ID card usage data, Ram is able to piece together complex behavioral patterns for both student groups and individuals.

For example, if student A, on multiple occasions, uses her ID card at the same location and time as student B, it stands to reason there is social interaction between the two. When extrapolated over time, detailed behavioral and social patterns emerge.

By tracking changes to these patterns over time, Ram has been able to accurately predict freshmen dropouts at an 85-90% rate, up from the university’s current success rate of 73% using traditional metrics.

The findings show promise for the use of machine learning methodologies and big data analytics in the AEC industry and real estate sector. For example, a similar approach could be applied to commercial office buildings, to identify tenants that are most at-risk for not renewing their lease.