DETERMINING THE INFLUENCE OF DATA ON WORKING WITH VIDEO MATERIALS ON THE ACCURACY OF STUDENT SUCCESS PREDICTION MODELS

Abstract

The object of this study is models for predicting students’ success, constructed on the basis of machine learning methods. The paper reports results of research into the problem of improving their accuracy by expanding the data set for training the specified models. The most available are data on student actions, which are automatically collected by learning management systems. Entering additional information about students’ work increases time and resources but allows the improvement of the accuracy of the models. In the study, information about students’ work with video materials, particularly the number and duration of views, was entered into the original data set. To automate the collection of this data, the plugin for the Moodle system has been developed, which stores information about user’s actions with the video player and the duration of watching video materials in the database. Model training was carried out using Naive Bayes (NB), logistic regression (LR), random forest (RF), and neural networks (NN) algorithms with and without video data. For the models using video viewing data, accuracy increased by 10%, balanced accuracy by 15%, and overall performance, expressed as area under the curve (AUC), increased by 14%. The highest prediction accuracy, with a difference of 1.8%, was obtained by models built using RF algorithms – 87.1% and NN – 85.3%. At the same time, the accuracy of the models obtained by the NB and LR algorithms was 70.7% and 76.5%. The increase in accuracy for them was 2.3% and 8.1%, respectively. Analysis of calculations confirms the assumption that students’ work with educational video materials is correlated with their success. The results make it possible to find a reasonable compromise between model development costs and its accuracy at the stage of data preparation for model training