Now-a-days, research in educational mining focuses on modelling student's performance. Many universities include large volumes of data related to student's details, performance, management details, educational process, and etc. Moreover, most of the data remains unused because inability of the university administration to handle it, also huge volumes of data are difficult to perform. In this paper, hybrid Feature Selection (FS) method namely Relief-F and Budget Tree-Random Forest (RFBT-RF) is proposed for selecting active features to reduce high dimensionality and handle uncertainty of data. The proposed feature selection method selects only relevant features instead of selecting redundant and irrelevant features for the classifiers. Also, RFBT-RF method is applied on multiple classifiers like Decision Tree (DT), Naive Bayes (NB) and Support Vector Machine (SVM) and K-Nearest Neighbour (KNN) for predicting the Student Academic Performance (SAP). RFBT-RF method was applied on three databases such as UCI (maths), UCI (Portuguese) and Collected dataset. Results showed that, RFBT-RF algorithm achieved 6.85% of improved SAP accuracy compare to the existing Logistic Regression (LR) model.

No abstract

The student academic prediction model helps to predict the student performance that helps the university to provide necessary care to the particular students. Efficient prediction model helps to encourage the student for better performance in the academic. In this research, the Relief-F Budget Tree Random Forest with Gray Wolf Optimization (RFBTRF-GWO) method is proposed for the feature selection. The Gray Wolf Optimization (GWO) helps to scale the relevant feature with ranking order from the features selected by the Relief-F Budget Tree Random Forest (RFBTRF). The selected features are given as input to the classifier for the effective prediction. The k-Nearest Neighbor (kNN) and Artificial Neural Network (ANN) are used for the classification. The proposed RFBTRF-GWO method is evaluated on the three datasets such as two UCI datasets and one collected dataset. The RFBTRF-GWO has a higher performance accuracy of 96.2 % while the existing method RFBTRF has an accuracy of 70.88 %.