Artificial Intelligence Flags Students at Risk of Dropping Out—and Triggers Personalized Support
A new artificial-intelligence system designed to identify students at risk of leaving school early has combined academic records, attendance patterns, family information and behavioral data to forecast performance and guide individualized support. The system, known as TC-Net, was tested using records from 395 students attending two schools in Portugal. Researchers report that the model explained about 60.57 percent of the variation in students’ final grades, while producing a mean squared error of 0.43 in performance prediction. The framework was then used to direct interventions toward students considered vulnerable: 90 percent of those identified as at risk engaged with at least one form of support, including tailored learning plans, tutoring and mentoring. The study presents the technology as an early-warning system for educators, although its results remain limited by the small and geographically narrow dataset.
Student dropout rarely results from a single event. Academic difficulties can interact with repeated absences, weak family support, social pressures, financial stress, mental-health problems and disengagement from school. By the time these difficulties become obvious to teachers, a student may already be far behind. The central idea behind TC-Net is to detect combinations of warning signals earlier, when support may still change the student’s trajectory. Rather than treating dropout as a simple yes-or-no outcome, the framework first predicts academic performance and then estimates whether a student’s profile resembles that of an at-risk learner. In principle, such a system could allow schools to move from reacting to failure toward offering assistance before failure becomes entrenched. The researchers emphasize, however, that the output is intended to support professional judgment, not to label students permanently or impose punitive decisions.
The underlying dataset, known as Student Performance or Student Final Grade Prediction, was obtained from the UCI Machine Learning Repository. It contains 33 attributes gathered from two Portuguese schools. These include grades from the first and second assessment periods, represented as G1 and G2, and the final grade, G3, along with study time, past failures and absences. The records also include age, sex, school, address type, family size and parental education. Social and behavioral variables cover internet access, romantic relationships, family relationships, free time and alcohol consumption during weekdays and weekends. Household characteristics include parental occupations, whether parents live together, educational support, paid classes and nursery attendance. The students had a mean age of 16.7 years, and their final grades averaged 10.4 out of 20. Absences ranged from zero to 75, with a median of four, while approximately 28 percent of the records were classified as at risk according to the study’s final-grade threshold.
Before training the model, the researchers applied several data-processing techniques intended to reduce noise and preserve useful patterns. Missing values were estimated using non-negative matrix factorization, or NMF, a method that represents a non-negative data table as the product of two smaller non-negative matrices. In educational data, one matrix can be interpreted as latent student characteristics and the other as patterns linking those characteristics to grades, attendance or social variables. Their combination can reconstruct incomplete entries without allowing implausible negative values. The team also used Local Outlier Factor, or LOF, to identify unusual records. LOF compares the local density around one observation with the density around its nearest neighbors; a student record lying in an unusually sparse region may indicate a data-entry error or an uncommon case requiring caution. Removing or separately examining such records can prevent extreme observations from dominating model training, although excluding unusual students can also erase precisely the cases that schools most need to understand.
Feature selection was performed with techniques including mutual information and LASSO, or least absolute shrinkage and selection operator. Mutual information measures how much knowing one variable reduces uncertainty about another, while LASSO adds a penalty to a regression model that pushes weak or redundant coefficients toward zero. Together, these methods were used to retain informative academic, behavioral and sociodemographic variables while reducing unnecessary inputs. The analysis found especially strong links between the final grade and earlier grades: Pearson correlation coefficients for G1 and G2 with G3 were reported as greater than 0.85. Absences showed a negative correlation with final performance, with a reported coefficient of −0.62. Weekday alcohol consumption and study time were also associated with performance patterns. These relationships do not prove that absences or alcohol use cause lower grades, because the variables may reflect broader circumstances such as health, family difficulties or disengagement.
The “T” and “C” in TC-Net refer to its two principal components: TabNet and Capsule Networks. TabNet is a neural architecture developed for structured, spreadsheet-like data. At successive decision steps, an attention mechanism assigns greater weight to selected features, allowing the model to focus on variables such as prior grades or attendance rather than treating every input as equally important. Its output can be converted through a sigmoid function into a probability between zero and one, representing the estimated likelihood that a student belongs to the at-risk category. This attention mechanism offers a degree of interpretability because it can indicate which features influenced a prediction. It does not, however, automatically establish causation, and a feature receiving high attention should not be treated as a direct explanation of a student’s difficulties.
The second component, a Capsule Network, was originally developed for computer-vision tasks in which the relationships among parts of an object matter. Instead of representing information as isolated scalar activations, capsule networks use groups of neurons that encode richer vectors and employ dynamic routing to determine which lower-level features should contribute to higher-level representations. In TC-Net, the capsules are used to model complex interactions among selected student features. A pattern involving attendance, previous grades and family circumstances might carry a different meaning from any of those variables alone. By routing information among capsules, the model attempts to preserve these higher-order relationships. The researchers argue that this gives TC-Net an advantage over systems that simply rank features independently, particularly when academic, behavioral and social factors combine in nonlinear ways.
The reported performance was stronger than that of the study’s baseline approach in several measures. TC-Net produced a mean squared error of 0.43, compared with 3.50 for the baseline model. Mean squared error is calculated by averaging the squared differences between predicted and actual values, so it penalizes large errors more heavily than smaller ones. The study also reports a minimum root mean squared error of 0.19 as additional features were included, although the exact evaluation conditions for that value are important when interpreting the comparison. The reported mean absolute error was 2.89, meaning that predictions differed from actual grades by about 2.89 grade units on average under the stated analysis. An R-squared value of 60.57 percent indicates that the model accounted for roughly three-fifths of the observed variance in final grades. That is a meaningful signal, but it also means that nearly 40 percent of the variation remained unexplained.
The researchers used five-fold cross-validation, dividing the data into five portions and repeatedly training on four while evaluating on the remaining portion. This approach helps estimate how a model performs on unseen records, particularly when datasets are small. Paired t-tests across the folds reportedly found statistically significant improvements over a Random Forest baseline for accuracy, precision, recall and F1 score at the conventional 0.05 threshold. A supplementary two-tailed Z-test produced a p-value of 0.06, which does not meet that threshold and therefore provides only marginal evidence of a difference under that test. The conflicting statistical signals illustrate why performance claims should be treated carefully when based on only 395 students. The study did not report confidence intervals or a detailed analysis of false positives and false negatives. Those omissions matter: incorrectly flagging a student could create stigma or waste scarce support resources, while failing to flag a struggling student could delay help.
Prediction was linked to a three-part intervention strategy. Students identified as vulnerable could receive tailored learning plans containing additional practice questions, video explanations, interactive modules or a slower pace through difficult material. Others could be referred to one-to-one tutoring, small-group study sessions, mentoring or peer support. The framework also allows for counseling, support groups, stress-management training and other mental-health services when social or emotional problems appear to be affecting schoolwork. According to the study, 90 percent of at-risk students engaged with an intervention, 80 percent participated in tailored learning plans and 70 percent accessed tutoring support. The researchers also report improvements in aggregated engagement measures after intervention, particularly attendance at tutoring and participation in mentorship activities. These findings suggest that data-guided outreach can connect students with assistance, but engagement is not the same as improved graduation or proven dropout prevention. The source material does not establish a randomized control group showing that TC-Net interventions directly reduced dropout rates.
A proposed real-time version of the system would connect TC-Net to a school’s learning-management platform. Risk scores could appear on a dashboard for teachers, counselors and academic advisers, while threshold alerts could notify staff when a student’s estimated risk rises. The system could recommend resources automatically, but educators would retain the ability to review or modify those recommendations. Feedback such as attendance at support sessions, new grades, learning-platform activity and behavioral signals could be returned to the model for later refinement. That arrangement could make the technology more responsive, but it also creates a substantial data-governance challenge. Student records can reveal sensitive information about family life, health, relationships and socioeconomic circumstances. The researchers describe anonymization, encryption, consent and fairness audits as safeguards, and stress that predictions should be used for support rather than punishment. Even with these protections, schools would need clear rules about who can access risk scores, how long data are retained and whether families can challenge an automated assessment.
The most important limitation is generalizability. TC-Net was developed from a relatively small, largely homogeneous sample drawn from only two Portuguese schools. Educational systems differ in grading policies, attendance rules, family structures, language, access to technology and definitions of dropout. A pattern that predicts difficulty in Portugal might perform differently in a rural school, a large urban district or a country with another curriculum. Capsule Networks also add technical complexity that may make the system difficult for educators to audit, despite TabNet’s partial interpretability. Future studies will need larger, multi-institutional datasets, external validation across countries and demographic groups, transparent error analysis and uncertainty estimates. Researchers should also test whether interventions improve long-term outcomes rather than merely increasing short-term engagement. AI may help schools notice patterns hidden in vast records, but its most defensible role is as a carefully monitored assistant—one that prompts human support while leaving the final understanding of a student’s circumstances to people who know how to listen.
Subject of Research: AI-based prediction of student academic performance and dropout risk, combined with personalized educational interventions.
Subject of Research: Technology and Engineering
Article Title: AI-Powered Student Dropout Prediction and Personalized Intervention Using TC-Net in Education
Article References: https://link.springer.com/article/10.1007/s44163-026-01241-z
Image Credits: AI Generated
DOI: 10.1007/s44163-026-01241-z
Keywords: artificial intelligence, machine learning, student dropout prediction, personalized intervention, TC-Net, TabNet, Capsule Networks, educational data mining, at-risk students, predictive analytics
Tags: academic and attendance data integrationacademic performance forecastingAI dropout predictionattendance and family impact on student successbehavioral data analysis in schoolsearly-warning system for schoolsearly-warning systems in educationeducational technology in Portugallimitations of AI in small datasetsmachine learning for student performancemachine learning in educationmental health and school engagementmental health and social factors in dropout predictionpersonalized educational interventionspredictive analytics in educationstudent behavioral data analysistailored learning plans and tutoringtailored learning plans for dropout preventiontargeted support for at-risk studentsTC-Net student risk assessmentTC-Net student risk model
