Improving Students’ Performance by Interpretable Explanations using Ensemble Tree-Based Approaches

Abstract

The careful analysis and evaluation of students’ results are an important part of the educational activity, with a potentially strong impact on the students’ future development. Seven classification algorithms, which are Decision Tree, Bagging, Random Forest, AdaBoost, Gradient Boosting, XGBoost, and LightGBM, were used in this research. In this paper, for our experiments we used two datasets, the first refers to classify and predict Portuguese language performance and the second for students’ level at courses. In this paper, we propose to identify the most appropriate classification technique to improve the prediction of students’ performance, interpreting it using the LIME algorithm. The obtained results using both datasets show that the model built using Decision Tree, outperforms the other constructed models. Our methodology consists of four major steps: i) analyzing and preprocessing the dataset; ii) optimizing the models using cross-validation and hyperparameter tuning; iii) comparing the performance of different ensemble tree-based models, and iv) interpreting the model by providing explanations. The development of explainable models can lead to important advantages: the model can be trusted, the transparency of the model helps to understand the underlying mechanisms that make the model work and opaque models can be interpreted without sacrificing their predictive performance.