In 2012, Georgia State University faced a critical problem: a diploma rate of 44%, among the lowest in the major American universities. Thousands of students abandoned each year for apparently small but insurmountable reasons for them, such as an unpaid bill or a missed pre-requisite course.
The Timothy Renick Provost chose a different path: no more general counseling or expensive tutoring, but a predictive system that identified first who risked abandoning, sending proactive alert to tutors at the right time.
The GPS Advertising system analysed 800 variables per student, from grades and attendance to interactions with support services. Thus, the university intervened weeks before the student stopped attending.
Six years later, the degree rate had risen to 55%, with 56,000 more graduates without significantly increasing the budget for support.
In 2026, the problem persists in Italy: 15-20% of students leave after the first year, with a dramatic regional gap between North (12%) and South (25%).Every abandonment represents a cost per family, university and country system.
The serious question is: can we predict abandonment *before it happens and intervene effectively? The answer is yes, without the need for artificial science fiction intelligence, but with essential data, reasonable statistical models and organizational will.
Real problem
ISTAT and Anvur data show that the drop-out rate varies significantly by geographical area and type of university, with estimated costs in hundreds of millions of euros per year. The North-South gap is alarming, with students in the South having twice the probability of leaving compared to the North-West.
Conceptual model
European and Italian studies identify powerful and counterintuitive predictors: the number of CFUs acquired in the first half of the year is the strongest sign, more than the maturity vote. The abandonment is the conclusion of a long process, traceable through progressive decommitment, visible in academic data and digital engagement.
Typical profile of the high-risk student includes middle first semester low, few CFU, low LMS frequency, choice of course as second option and low maturity grade.
Strict formalisation
Data collection and integration
The data are often fragmented into silos: secretariat, LMS, orientation, attendance. The first step is to build a pipeline that integrates these sources into a unique dataset for students.
Feature engineering
The most predictive variables include CFU first semester, excess rate exams, average grades, days of LMS inactivity, grade maturity, course choice, commuting status and work.
Predictive model
Random Forest and Gradient Boosting reach AUC between 0.82 and 0.88, effectively distinguishing risky students from non-risky ones. The choice of alert risk threshold depends on the resources available: higher thresholds reduce false positives but intercept fewer students, lower thresholds increase coverage at the cost of more false alarms.
Interpretability
The first semester CFUs dominate the importance of the features, confirming the centrality of the real academic behavior in the first months.
Example or case study
Georgia State University and other universities (Polytechnic University of Milan, University of Trento, Purdue University) have demonstrated that early warning systems with proactive interventions significantly reduce the rate of abandonment, with success rates of interventions up to 72%.
Lab / exercise
Basic level: extract raw data from management and identify students with zero CFU in February.
** Intermediate level:** Build a simple logistic regression model with the main features and evaluate the AUC.
Level research-grade: implement a Random Forest or Gradient Boosting model with cross-validation and risk threshold optimization.
Datasets and recommended materials: anonymous datasets of Italian students, examples of SQL queries of integration, Python scripts for feature engineering and modeling.
Typical error to avoid
Do not act only on the basis of the model: the value is in the prompt and proactive intervention. Also avoid communicating “risk” labels to students without a positive framing, not to generate self-fulfilling prophecies or demoralization.
Quiz or checkpoint
- What is the most useful predictor of university drop-out?, Why is it important to choose the risk threshold based on resources?, What are the ethical risks in using predictive models for drop-out?, How can you monitor the equity of the model?
The point predicting university abandonment is a discipline of decision under uncertainty that combines data, models and organizational action. data exist, technology is accessible, but it needs a culture that values proactive intervention and measures impact. the challenge is not technical, but cultural and ethical: using data to really help the student, not to monitor him.
A well-implemented predictive system can reduce the drop-out rate by 2-4 percentage points in 18 months, with a return on investment in the first year. It is a discipline that every Italian university should adopt to enhance human capital and improve the efficiency of the university system.