Modeling Exercise Relationships in E-Learning: A Unified Approach.

In an e-learning system, relationships between a large amount of exercises are complex and multi-dimensional; measuring the relationships and arranging curriculums accordingly used to be time consuming and costly tasks which require either enormous log collection or large-scale human annotations. Moreover, accurately quantifying the relationships is difficult because there are too many factors which affect our measurement based on the data, such as the ability of exercise takers and the subject bias of annotators. To overcome these challenges, we propose a unified model that extracts information from both human annotations and usage log using regression analysis. The proposed model is applied to quantify the similarity, difficulty, and prerequisite relationships between every two exercises in a curriculum. As a case study, we collaborate with Junyi Academy, a popular e-learning platform similar to Khan Academy, and infer the pairwise relationships of 370 exercises in its mathematics curriculum. We show that the model can predict exercise relationships as well as an expert does with human annotations of a few sample exercise pairs (2% in our experiments). We expect the introduction of the proposed unified model can improve the relationships among exercises and learning pathways of students in other e-learning platforms.