Heterogeneous Features Integration in Deep Knowledge Tracing

Abstract

Knowledge tracing is a significant research topic in educational data mining. The goal is to automatically trace students’ knowledge states by analyzing their exercise performance. Recently proposed Deep Knowledge Tracing (DKT) model has shown a significant improvement to solve this task by applying deep recurrent neural networks to learn interaction between knowledge components and exercises. The input of the model is only the one-hot encoding to represent the exercise tags and it excludes all other heterogeneous features, which may degrade the performance. To further improve the model performance, researchers have analyzed the heterogeneous features and provided manual ways to select the features and discretize them appropriately. However, the feature engineering efforts are not feasible for data with a huge number of features. To tackle with them, we propose an automatic and intelligent approach to integrate the heterogeneous features into the DKT model. More specifically, we encode the predicted response and the true response into binary bits and combine them with the original one-hot encoding feature as the input to a Long Short Term Memory (LSTM) model, where the predicted response is learned via Classification And Regression Trees (CART) on the heterogeneous features. The predicted response plays the role of determining whether a student will answer the exercise correctly, which can relieve the effect of exceptional samples. Our empirical evaluation on two educational datasets verifies the effectiveness of our proposal.