Achieving optimal trade-off for student dropout prediction with multi-objective reinforcement learning

Student dropout prediction (SDP) in educational research has gained prominence for its role in analyzing student learning behaviors through time series models. Traditional methods often focus singularly on either prediction accuracy or earliness, leading to sub-optimal interventions for at-risk stud...

Full description

Saved in:
Bibliographic Details
Published in:PeerJ. Computer science 2024-04, Vol.10, p.e2034-e2034, Article e2034
Main Authors: Pan, Feng, Zhang, Hanfei, Li, Xuebao, Zhang, Moyu, Ji, Yang
Format: Article
Language:English
Subjects:
Analysis
Artificial Intelligence
Computer Education
Data Mining and Machine Learning
Data Science
Discovery and exploration
Envelope Q-learning
Markov processes
Multi-objective reinforcement learning
Neural Networks
Outer space
Student dropout prediction
Trade-off
Vector reward
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Student dropout prediction (SDP) in educational research has gained prominence for its role in analyzing student learning behaviors through time series models. Traditional methods often focus singularly on either prediction accuracy or earliness, leading to sub-optimal interventions for at-risk students. This issue underlines the necessity for methods that effectively manage the trade-off between accuracy and earliness. Recognizing the limitations of existing methods, this study introduces a novel approach leveraging multi-objective reinforcement learning (MORL) to optimize the trade-off between prediction accuracy and earliness in SDP tasks. By framing SDP as a partial sequence classification problem, we model it through a multiple-objective Markov decision process (MOMDP), incorporating a vectorized reward function that maintains the distinctiveness of each objective, thereby preventing information loss and enabling more nuanced optimization strategies. Furthermore, we introduce an advanced envelope Q-learning technique to foster a comprehensive exploration of the solution space, aiming to identify Pareto-optimal strategies that accommodate a broader spectrum of preferences. The efficacy of our model has been rigorously validated through comprehensive evaluations on real-world MOOC datasets. These evaluations have demonstrated our model's superiority, outperforming existing methods in achieving optimal trade-off between accuracy and earliness, thus marking a significant advancement in the field of SDP.
ISSN:2376-5992
2376-5992
DOI:10.7717/peerj-cs.2034