The prevalence of shallow reading in online digital learning is steadily increasing, which has sparked interest in revealing the mechanisms behind shallow reading behavior, especially analyzing the causal relationship between its constituent features and learning performance. However, current causal analysis methods have many limitations in terms of experimental conditions, data independence assumptions, and analysis costs. Drawing on the application experience of Markov chain theory in the field of causality, this study adopts the structure-agnostic model (SAM) algorithm to design the structure, parameter loss, and learning process, and proposes an evaluation method for causal exploration based on generative adversarial neural networks (GANs). The study shows that the proposed maximum mean diversity (MMD) optimization method improves the stability of the model analysis results and clarifies that reading speed is a key factor in the occurrence of shallow reading behavior.

Reading rate is the number of words a user reads per unit of time, also known as reading speed. As online learning has been continuously growing, research on tracking learning behavior is undertaken, and the reading rate was beginning to be applied to measure the learning process. However, current calculating methods are based on the traditional learning environment of a paper-based course, and further research is needed to determine whether they can be used to track students’ learning in an online learning environment and integrate with their online reading habits. To this end, this study proposed a new method based on online learning used to calculate reading rate as a measure to track students’ learning status, after analyzing students’ reading behavior in the e-book learning system. This method not only can identify students with learning difficulties through reading rate outliers, but also provide urgent feedback to students on their reading status.

As a kind of data that can reflect learning status, e-book logs have been widely used in learning analytics, especially for the prediction of academic performance. However, the best prediction model cannot be found without determining the contribution of e-book logs to the prediction performance of the model and its creation process. To this end, this study used the scikit-learn, a free software machine learning library, to analyze learning performance of 234 participants by learning behavior logs, which were collected by an e-book system. Finally, six prediction models containing Decision Tree, Random Forests, XGBoost, Logistic Regression, Support Vector Machines, and K-nearest Neighbors were created. Also, the contribution of e-book logs on the establishment of different prediction models was obtained by three feature importance calculation methods, i.e., the impurity-based feature importance, coefficients feature importance, and permutation feature importance. Based on statistical results, it was concluded that the Decision Tree and Random Forests had the best prediction performance, which was compared to the other four models, with prediction performance scores ranging from 0.7 to 0.8. Besides, the four data features of Prev, Highlight, Maker, and Next were found to have the greatest impact on model prediction creation.