Predição de desempenho de alunos no próximo período baseado no histórico de notas

Abstract

Contrary to the gradual growth in the number of undergraduate enrollments, there is the problem of student evasion and retention. According to data from the 2018 Higher Education Census by INEP, only 37.9% of students entering in 2010 completed their studies and 5.3% were continuing their studies after 6 years (INEP, 2018). In this sense, the ability to predict student performance in subjects is fundamental to help in the educational decisions of Higher Education Institutions (HEI). The main objective of the present study is to predict student performance using Neural Collaborative Filtering (NCF) in addition to comparing this model with classical models of Collaborative Filtering (CF). Students will be classified as passing or failing. Templates generate recommendations for students based on grade history. The models were submitted to non-sensitive data from the Federal University of Ceará (FUC) Crateús campus. As a metric to analyze the performance of the models and compare them, the average of the metrics was used: accuracy , precision , recall and F1_score. The tested models had good results and among them the classic model Generalized Matrix Factorization (GMF) had the best performance indicating Accuracy of 0.891, Precision of 0.931, Recall of 0.871 and F1_score of 0.867 based on the data used for training, testing and validating the models. Finally, the Multi-Layer Perceptron (MLP) model had the following results: Accuracy of 0.841, Precision of 0.841, Recall of 0.913 and F1_score of 0.840. The Neural Matrix Factorization (NMF) model, in turn, had the following results: Accuracy of 0.863, Precision of 0.896, Recall of 0.870 and F1_score of 0.848.