The main goal of learning analytics and early detection systems is to extract knowledge from student data to understand students' trends of activities towards success and risk and, therefore, design intervention methods to improve learning performance and experience. However, many factors contribute to the challenge of designing and building effective learning analytics systems. Because of the complexity of heterogeneous student data, models designed to analyze it frequently neglect temporal correlations in the interest of convenience. Moreover, the performance descriptions gained from the student data model or prediction results from the analytical models do not always help explain the "why" and "how" behind it. Furthermore, domain specialists are unable to participate in the knowledge discovery process since it necessitates significant data science abilities, and an analytical model is a black box to them.
This research aims to develop analytical models that enable domain experts to study their students' performance behavior and explore trustworthy sources of information with the help of explanations on the analytics. Our work demonstrates various approaches to using the temporal aspect of heterogeneous student data to build analytical models: weighted network analysis, unsupervised cluster analysis, and recurrent neural network analytics. The description, implementation process, and findings of each method are presented as technical contributions to the temporal analysis of student data. We experiment with all these analytical models that highlight the complexity of heterogeneous-temporal data, model building, decision-making tasks, and the need for a more in-depth focus on visual information of analytics with state-of-art explainable AI tools and techniques.
Our work underscores a need for developing a robust way to integrate the possibilities inherent within each approach. To achieve this goal, we present a comprehensive yet flexible and empirical framework to support the design and development of analytical models to extract meaningful insights about students' academic performance and identify early actionable interventions to improve the learning experience. We illustrate our framework on three applications (e.g., student network model, unsupervised clustering model, and recurrent neural network analytics) to demonstrate the value of this framework in addressing the challenges of using student data for learning analytics. These applications present vast opportunities to benefit students' learning experience by implementing flexible educational data representations, fitting different predictive models, and extracting insights for designing prescriptive analytics and building strategies to overcome perceived limitations.
An academic institution's culture drives its ability to accept, leverage, and deploy predictive and prescriptive analytics to enhance the workflow of maximizing pedagogical outcomes. We believe that our work will aid in the future development or refinement of a set of design standards for learning analytics systems.