Um framework apoiado por inteligência artificial para análise de aprendizagem autorregulada de estudantes em ambientes virtuais de aprendizagem

Abstract

The advancement of digital technologies has transformed teaching and learning processes, making Virtual Learning Environments and Intelligent Learning Environments central spaces for investigating student behavior. These environments generate large volumes of interaction data that enable the analysis of how students plan, monitor, and evaluate their study strategies, which are fundamental dimensions of self-regulated learning. However, transforming these records into valid, interpretable, and pedagogically meaningful indicators still represents a challenge for educational research. This thesis proposes and validates the EDM Framework, a methodological model grounded in Educational Data Mining, developed to identify and analyze self-regulated learning behaviors based on interaction data. The framework systematically organizes the stages of data collection, preprocessing, mining, and interpretation, integrating principles from educational psychology, with emphasis on Zimmerman’s self-regulated learning model, and unsupervised clustering techniques. Throughout the research, the K-Means, HDBSCAN, and Agglomerative Clustering algorithms were tested and compared across four articulated case studies, which evolved from a proof of concept to practical application in a real educational context. In the final case study, conducted in a real “Introduction to Python Language” course, both hierarchical (Agglomerative) and non-hierarchical (K-Means) clustering approaches were applied using a parsimonious two-cluster solution to detect low and high self-regulated learning profiles over the duration of the course. The results showed that students with higher levels of self-regulated learning exhibited more consistent engagement patterns with educational resources and achieved better academic performance. Statistically significant differences between the identified profiles were observed at multiple points throughout the course, reinforcing the consistency of the proposed approach. Overall, the findings confirm that it is possible to identify consistent evidence of self-regulated learning from interaction data in virtual learning environments, demonstrating the effectiveness of clustering algorithms in forming profiles aligned with the theoretical dimensions of self-regulated learning. This thesis contributes to the advancement of the state of the art by offering a methodological process that integrates theoretical foundations and computational procedures, supporting the development of more adaptive, personalized educational systems that are sensitive to students’ self-regulatory strategies.