Feature Mining Algorithm for Student Academic Prediction Based on Interpretable Deep Neural Network
Yiming Luo, Zhuo Wang
Abstract
Predicting student academic performance is one of the pivotal issues of concern in the educational domain. With their outstanding predictive capabilities, neural network algorithms are widely used for predicting student academic performance. This paper primarily focuses on the crucial issue of the interpretability of deep neural network models in the field of educational data mining and proposes an interpretable deep learning model, namely Log-LassoNet, which is applied to educational data. This paper conducts modelling analysis on the China Education Survey data set by introducing noise as additional fake features. The model demonstrates satisfactory predictive and feature selection capabilities, outperforming current mainstream algorithms such as Lasso regression and random forests. Moreover, through feature mining, it was found that the expected and health-related features play an important role in forming high-level features that impact student academic performance, which provides a foundation for exploring the factors affecting students' academic performance.