Grade Prediction From Multi-Valued Click-Stream Traces via Bayesian-Regularized Deep Neural Networks

Abstract

Learning activities that have been designed linearly in small private online courses and thematic massive open online courses often result in learners undergoing similar learning journeys. While most learners perform activities leading to similar click-stream sequences, they achieve contrasting grades. This phenomenon of similar online learning behavior resulting in different learning outcomes is termed as the multi-valued (MV) problem. Existing grade prediction models are not designed to handle MV inputs and this problem is further aggravated because the MV characteristic cannot be expressed analytically due to the complexity of these sequential inputs. In this work, the detrimental impact of MV inputs on grade prediction is highlighted. A new MV-sensitive grade prediction model is proposed to quantify the MV characteristic of the dataset. As opposed to existing deep neural network based grade prediction models, the proposed algorithm is able to capture the uncertainty associated with each click-stream sequence. This results in the model organizing its feature space to discriminate the MV inputs. With the newly formulated surrogate measures, and by removing these identified MV inputs, experiment results show that the proposed models achieve lower prediction error compared to existing algorithms when evaluated on two online-learning datasets.