Academic achievement inspired machine learning methods for student grade prediction and at-risk detection

Grade prediction is one of the several sub-disciplines among learning analytics transcends and has received increasing attention in recent years and remains one of the most challenging tasks. Grade prediction plays a central role in the development of data-informed approaches for at-risk student det...

Full description

Saved in:
Bibliographic Details
Main Author: Wei, Qiu
Other Authors: Andy Khong W H
Format: Thesis-Doctor of Philosophy
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/175913
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Grade prediction is one of the several sub-disciplines among learning analytics transcends and has received increasing attention in recent years and remains one of the most challenging tasks. Grade prediction plays a central role in the development of data-informed approaches for at-risk student detection and early intervention. This thesis addresses the grade prediction and at-risk student detection problem from student prior grade data. For grade prediction and at-risk student detection, this thesis focuses on the modeling of performance-centric data. Although several types of data exists for prediction, performance-centric data such as assessment scores are often made accessible to administrators, policymakers, instructors, and student care support personnel involved in administering interventions. To this end, sustained efforts have been focused on leveraging prior grades attained in consecutive (past) semesters to predict grades of pilot courses that have been registered in the upcoming or current semester. To achieve a high detection rate of at-risk student and reduce false alarm, a dual-mode long short-term memory (LSTM) model is proposed. As will be shown in this thesis, the proposed two-stage architecture employs the weighted-loss function and the short-term gated LSTM. Performance of the proposed grade prediction performance is evaluated on the dataset from three departments in a university. The proposed model improves the performance of LSTM by 28.8\% in terms of F1 score for at-risk classification. While the LSTM model is able to model the temporal information in student prior grades, there are other aspects of student performance that are important for grade prediction. To jointly consider the three important aspects\textemdash temporal dynamics of prior courses across previous semesters, short-term performance consistency, and relative performance against peers, an academic achievement inspired grade prediction (A2GP) model is proposed. Results obtained highlighted that the A2GP model improves the performance of LSTM and graph convolutional network (GCN) by 19.0\% and 63.3\%, respectively, in terms of F1 score for at-risk classification. Existing grade prediction methods do not take into account the constraints imposed on grade vectors. As a third contribution, a grade prediction framework is proposed for efficient information encoding that includes a relative performance module, a logic reasoning module, and a grade prediction module. The relative performance module employs a cohort-grade distribution that incorporates relative performance and the logic reasoning module embedded the sparse constraints as the representation. With the Transformer within the grade prediction module that encodes all relevant information, the model is able to outperform all baseline models for at-risk student detection.