Uso da Inteligência Artificial (IA) para predição da conclusão do ensino superior por alunos com deficiência

Abstract

The enrollment of people with disabilities in higher education has been growing and poses many challenges for universities. The development of support actions suited to the needs of students with disabilities in higher education with a view to completing the course is essential in order to provide students with a new occupational perspective, making them fit for the job market. Prediction studies, through artificial intelligence, in the educational context for this public can contribute to improve inclusion policies and minimize social, economic, political, academic and financial losses caused to all those involved in the educational process. The objective of this study was to develop and evaluate the performance of an algorithm model in predicting higher education completion by students with disabilities at a public university in Minas Gerais. This is an exploratory, quantitative and retrospective study based on a database made available by the institution, comprising 563 students with disabilities enrolled in undergraduate courses between 2001 and 2020. Of sociodemographic and academic variables indicated at the time of admission to the institution. Five algorithm models were tested: logistic regression, K-Nearest Neighbors (KNN); Light Gradient Boosting Machine (LGBM); Randon Forest (RF) and Extreme Gradient Boosting (XGBoost). In order to classify and identify the relationship between the input variables and the outcome variable, the post hoc SHAP method of interpretation was used. The results indicated that the model with the best performance for prediction was XGBoost, with average accuracy (ACC=76.38%); F1-Scoore of 76.54% and ROC curve equal to 80.71%. The main predictive factors for completing the course for this public were the type of high school, form of admission, grade in the selection process, age and gender. There is a dynamic interaction between environmental, contextual and personal factors, with the first two being the most relevant. The results in this sample indicate that aI can be used with moderate accuracy to early identification of predictive factors for non completion of the course by students with disabilities and can be an important tool in order to help teachers and the institution to direct actions and resources according to the needs of students with disabilities, contributing to their permanence and participation in this context with consequent conclusion of the course and future occupational perspective.