Deep Learning Models and Interpretations for Multivariate Discrete-Valued Event Sequence Prediction

We propose an embedding-based deep learning model architecture for raw clickstream event sequences, which has special characteristics, such as being multivariate discrete-valued. We evaluate the proposed architecture on a Stanford University MOOC dataset, which consists of clickstream-level raw log event data collected during student sessions in the MOOC. We introduce empirical results achieved by various configurations of the architecture on the student final grade regression task. Apart from the regression experiments, we also propose three visual interpretation techniques for explaining the black-box Temporal Convolutional Neural Network and Recurrent Neural Networks models. The goal is to provide easily applicable interpretations which can be used by domain experts without any Machine Learning technical expertise. Based on the visual interpretations, we were able to identify student behavior patterns from raw data, in line with educational research literature.