Machine Learning Applications for Early Warning Systems in Education

MUSSO, M. F.

Oxford

Conferencia; International Conference on Digital Image and Signal Processing (DISP?19).; 2019

Corgascience Limited

After successful implementations of machine-learning methods based on neural network models in national security applications for the prediction of terrorist activities and ongoing work in the medical field, we have focused our efforts on several projects in the field of education and educational assessment. In an international large-scale project in primary education, we developed precise models estimating the combined impact of all school and teacher characteristics as well as background and non-school factors on student learning. The goal was to develop predictive models of academic performance in mathematics and language, establishing the main commonalities and differences between them, so as to be able to determine the most significant contributors to the differences in performance, and therefore suggest possible interventions on the low achievement groups that could impact positively in their performance. Later, we have examined the full range of learning outcomes in higher education, looking at the learning trajectory of the students and developing predictive models for some key outcomes. Thus, we worked on predicting and understanding different key points in a student?s academic trajectory such as grade point average (GPA), academic retention and degree completion, which could allow targeted intervention programs in higher education. With the implementation of some new modeling techniques, it was possible to identify the level of participation of each variable involved in the modeling of the problem, while achieving great accuracy in the predictive classifications, at levels of precision not usually achieved by traditional approaches. In this project we carried out three studies which addressed the prediction of: 1) GPA at the end of the second academic year; 2) drop-outs at the end of the second academic year; and 3) degree completion within a 5-year period. Multilayer perceptron ANN models with a backpropagation algorithm and with various activation functions were built to predict the various outcomes. Results demonstrated a very high accuracy of the ANN in the classification of students that belonged to each educational outcome group (GPA, drop-outs, and those that achieved degree completion). In addition, ANN provided information on those predictors that best modeled each group. The precise and early identification of those students at risk of dropping out from their university studies, as well as those that take longer than 5 years in completing the university program, allows more targeted interventions which can positively influence educational outcomes, raising academic perseverance and increasing degree completion rates. Finally, I will present a vision of the implementation of an ?intelligent classroom? in which with appropriate assessment methods and fully integrated use of multimodal data through machine-learning applications, continuous assessment with no testing could be implemented, providing teachers and other stakeholders with on-going data on student learning in a natural setting. The technical aspects of the models used, as well as the implementation of early-warning detection systems in education, using machine-learning approaches will be discussed.