Detecting Students At-Risk Using Learning Analytics

Abstract

The issue of supporting struggling tertiary students has been a long-standing concern in academia. Universities are increasingly devoting resources to supporting underperforming students, to enhance each student’s ability to achieve better academic performance, alongside boosting retention rates. However, identifying such students represents a heavy workload for educators, given the significant increases in tertiary student numbers over the past decade. Utilising the power of learning analytic approaches can help to address this problem by analysing diverse students' characteristics in order to identify underperforming students. Automated, early detection of students who are at potential risk of failing or dropping out of academic courses enhances the lecturers' capacity to supply timely and proactive interventions with minimal effort, and thereby ultimately improve university outcomes. This thesis focuses on the early detection of struggling students in blended learning settings, based on their online learning activities. Online learning data were used to extract a wide range of online learning characteristics using diverse quantitative, social and qualitative analysis approaches, including developing an automated mechanism to weight sentiments expressed in post messages, using combinations of adverbs, strengths. The extracted variables are used to predict academic performance in timely manner. The particular interest of this thesis is on providing accurate, early predictions of students’ academic risk. Hence, we proposed a novel Grey Zone design to enhance the quality of binary predictive instruments, where the experimental results illustrate its positive overall impact on the predictive models, performances. The experimental results indicate that utilising the Grey Zone design improves prediction-accuracy by up to 25 percent when compared with other commonly-used prediction strategies. Furthermore, this thesis involves developing an exemplar multi-course early warning framework for academically at-risk students on a weekly basis. The predictive framework relies on online learning characteristics to detect struggling students, from which was developed the Grey Zone design. In addition, the multi-course framework was evaluated using a set of unseen datasets drawn from four diverse courses (N = 319) to determine its performance in a real-life situation, alongside identifying the optimal time to start the student interventions. The experimental results show the framework’s ability to provide early, quality predictions, where it achieved over 0.92 AUC points across most of the evaluated courses. The framework's predictivity analysis indicates that week 3 is the optimal week to establish support interventions. Moreover, within this thesis, an adaptive framework and algorithms were developed to allow the underlying predictive instrument to cope with any changes that may occur due to dynamic changes in the prediction concept. The adaptive framework and algorithms are designed to be applied with a predictive instrument developed for the multi-course framework. The developed adaptive strategy was evaluated over two adaptive scenarios, with and without utilising a forgetting mechanism for historical instances. The results show the ability of the proposed adaptive strategy to enhance the performance of updated predictive instruments when compared with the performance of an unupdated, static baseline model. Utilising a forgetting mechanism for historical data instances led the system to achieve significantly faster and better adaptation outcomes.