Os efeitos de usar estimativas de conhecimento do aluno em programação de computadores em modelos livres de sensores de detecção da emoção confusão

Abstract

The research area of Affective Computing has been looking for ways to improve the detection of student confusion in computer-based learning environments. Environments capable of detecting student confusion can use different pedagogical strategies, such as intervening in the environment and helping students resolve their confusion or controlling it to benefit their learning. The author is interested in contributing to state the art in detecting confusion without using physical sensors (sensor-free) in the context of computer programming learning. The Thesis hypothesized that using data on student knowledge estimates and data on student interaction with the computer-based learning environment can improve the performance of sensor-free machine learning models in detecting student confusion in tasks of programming learning compared to baseline models. Baseline models represent related works, which developed their models using only student environment interaction data. The Thesis hypothesis is justified in cognitive theories of emotions, which relate the confusion with appraisals of incompatibility between the information that comes to the student and the student's mental model, such as the mental model of prior knowledge. To verify its hypothesis, the Thesis generated several machine learning models that represent the Thesis approach (Thesis hypothesis) and the baseline approach (related works) for different configurations of observation time windows (5, 10, 20, 40, 60, 90, 120, 180, 240 and 360 seconds and variable) and different algorithms. Statistical tests compared the results of each approach (Thesis and baseline). Methods were also applied to verify the models' most relevant data and the generalization performance for students with heterogeneous characteristics. The machine learning models were trained and tested with samples formed by data collected from 62 technical and higher education students for five months while solving exercises in a programming software adapted for the Thesis. Statistical tests showed that the best models of the Thesis approach presented superior and significant predictive accuracy compared to the best baseline models in all observation windows. In a list of the ten most relevant data attributes for the best models of the Thesis approach, five were attributes about interaction with the environment, and the other five were attributes about estimates of student knowledge. Regarding the performance of generalization for students with heterogeneous characteristics, significant differences were found between the approaches only in observation windows of 5, 10, and 20 seconds. In these windows, the performance of the thesis approach's best models was superior to that of the best baseline models. The results presented positive evidence that supports the hypothesis raised that estimates of student knowledge can improve the performance of sensor-free confusion detection models in computer programming tasks. The Thesis presents discussions for several other intermediate results and the scenarios where the Thesis approach is most advantageous.