Supervised learning through k-nearest neighbor, used in the prediction of university teaching performance

Abstract

This study initially seeks to identify the most optimal supervised learning algorithm to be used in predicting the perception of teacher performance, and then to evaluate its performance indicators that validate its predictive capacity. For this, the Matlab R2021a software is used; the experimental results determine that the supervised learning algorithm K-Nearest Neighbor Weighted (Weighted KNN) will be correct in 98.10% in predicting the perception of teaching performance, this has been validated by carrying out two evaluations through its performance indicators obtained in the confusion matrix and the receiver operating characteristic (ROC) curve, in the first evaluation an average sensitivity of 97.9%, a specificity of 99.1%, an accuracy of 98.8% and a precision of 96.7% are observed, thus validating the ability of the Weighted KNN model to correctly predict the perception of teacher performance; while in the receiver operating characteristic (ROC) curve, values of the area under the curve (AUC) equal to 0.99 and 1 are obtained, with this it is possible to validate the capacity that the model will have to distinguish between the 4 classes of the perception of the university teaching performance.