Academic Prediction in Multi-modal Learning Environments Using Data Fusion

In this paper, we present a proposal to predict the academic performance of university students in multimodal and blended learning environments based on data collected from different sources. To achieve this goal, we combined and processed data from 135 students and different variables from four different sources. First, we combined and preprocessed the data to create a summary dataset in numerical and categorical format. Then, we used different white-box classification algorithms provided by the data mining tool Weka to select the best algorithm. We found that the PART algorithm showed the best performance on the quality metrics, with a ROC range of 0.917. To further improve our prediction, we applied attribute selection algorithms, with ClassifierSubsetEval performing best with J48, with a ROC range of 0.9380. In addition, we used two machine learning algorithms, voting and stacking, and found that the best result was obtained with the Jrip algorithm and the voting method, with a ROC range of 0.9330. Finally, we presented our best predictive model, which is a hybrid of classification and machine learning algorithms, with a ROC range of 0.9420. We believe that this model can help faculty take corrective action for students who are at risk of dropping out or failing.