Contemporary education is a vast field that is concerned with the performance of education systems. In a formal e-learning context, student dropout is considered one of the main problems and has received much attention from the learning analytics research community, which has reported several approaches to the development of models for the early prediction of at-risk students. However, maximizing the results obtained by predictions is a considerable challenge. In this work, we developed a solution using only students’ interactions with the virtual learning environment and its derivative features for early predict at-risk students in a Brazilian distance technical high school course that is 103 weeks in duration. To maximize results, we developed an elitist genetic algorithm based on Darwin’s theory of natural selection for hyperparameter tuning. With the application of the proposed technique, we predicted the student at risk with an Area Under the Receiver Operating Characteristic Curve (AUROC) above 0.75 in the initial weeks of a course. The results demonstrate the viability of applying interaction count and derivative features to generate prediction models in contexts where access to demographic data is restricted. The application of a genetic algorithm to the tuning of hyperparameters classifiers can increase their performance in comparison with other techniques.

This paper describes a nationwide learning analytics initiative in Uruguay focused on the future implementation of governmental policies to mitigate student retention and dropouts in secondary education. For this, data from a total of 258,440 students were used to generate automated models to predict students at risk of failure or dropping out. Data were collected from primary and secondary education from different sources and for the period between 2015 and 2020. Such data contains demographic information about the students and their trajectories from the first grade of primary school to the second grade of secondary school (e.g., student assessments in different subjects over the years, the amount of absences, participation in social welfare programs, and the zone of the school, among other factors). Predictive models using the random forest algorithm were trained, and their performances were evaluated with F1-Macro and AUROC measures. The models were planned to be applied in different periods of the school year for the regular secondary school and for the technical secondary school ((before the beginning of the school year and after the first evaluation meeting for each grade). A total of eight predictive models were developed considering this temporal approach, and after an analysis of bias considering three protected attributes (gender, school zone, and social welfare program participation), seven of them were approved to be used for prediction. The models achieved outstanding performances according to the literature, with an AUROC higher than 0.90 and F1-Macro higher than 0.88. This paper describes in depth the characteristics of the data gathered, the specifics of data preprocessing, and the methodology followed for model generation and bias analysis, together with the architecture developed for the deployment of the predictive models. Among other findings, the results of the paper corroborate the importance given in the literature of using the previous performances of the students in order to predict their future performances.

This paper describes the application of Data Science and Educational Data Mining techniques to data from 4529 students, seeking to identify behavior patterns and generate early predictive models at the Universidad de la República del Uruguay. The paper describes the use of data from different sources (a Virtual Learning Environment, survey, and academic system) to generate predictive models and discover the most impactful variables linked to student success. The combination of different data sources demonstrated a high predictive power, achieving prediction rates with outstanding discrimination at the fourth week of a course. The analysis showed that students with more interactions inside the Virtual Learning Environment tended to have more success in their disciplines. The results also revealed some relevant attributes that influenced the students’ success, such as the number of subjects the student was enrolled in, the students’ mother’s education, and the students’ neighborhood. From the results emerged some institutional policies, such as the allocation of computational resources for the Virtual Learning Environment infrastructure and its widespread use, the development of tools for following the trajectory of students, and the detection of students at-risk of failure. The construction of an interdisciplinary exchange bridge between sociology, education, and data science is also a significant contribution to the academic community that may help in constructing university educational policies.

Predicting in advance the likelihood of students failing a course or withdrawing from a degree program has emerged as one of the widely embraced applications of Learning Analytics. While the literature extensively addresses the identification of at-risk students, it often doesn’t evolve into actual interventions, focusing more on reporting experimental outcomes than on translating them into real-world impact. The goal of early identification is straightforward, empowering educators to intervene before actual failure or dropout, but not enough attention is paid to what happens after the students are flagged as at risk. Interventions like personalized feedback, automated alerts, and targeted support can be game-changers, reducing failure and dropout rates. However, as this paper shows, few studies actually dig into the effectiveness of these strategies or measure their impact on student outcomes. Even more striking is the lack of research targeting stakeholders beyond students, like educators, administrators, and curriculum designers, who play a key role in driving meaningful interventions. The paper explores recent literature on automated academic risk prediction, focusing on interventions in selected papers. Our findings highlight that only about 14% of studies propose actionable interventions, and even fewer implement them. Despite these challenges, we can see that a global momentum is building around Learning Analytics, and institutions are starting to tap into the potential of these tools. However, academic databases, loaded with valuable insights, remain massively underused. To move the field forward, we propose actionable strategies, like developing intervention frameworks that engage multiple stakeholders, creating standardized metrics for measuring success and expanding data sources to include both traditional academic systems and alternative datasets. By tackling these issues, this paper doesn’t just highlight what is missing; it offers a roadmap for researchers and practitioners alike, aiming to close the gap between prediction and action. It’s time to go beyond identifying risks and start making a real difference where it matters most.