Predicting Student Performance Using a Cost-Sensitive Deep Learning Network With Sparse Attention Mechanism

doi:10.1109/TE.2025.3637343

Predicting Student Performance Using a Cost-Sensitive Deep Learning Network With Sparse Attention Mechanism

2026-01-01

0

PRE

AI

DOI：10.1109/TE.2025.3637343

Original

Original request for help

Save

Abstract

En 中文

Contribution: This study proposes a student performance prediction model that leverages real-world data on students' daily academic activities to forecast their academic outcomes. The predicted results can be used to provide students with personalized and adaptive support. Background: Traditional educational assessment systems face systemic challenges such as weak data support, high subjectivity, lack of scientific rigor, and insufficient personalized guidance. As a result, these assessments struggle to objectively, comprehensively, and accurately reflect the learning process and outcomes. In contrast, artificial intelligence (AI) technologies have demonstrated significant potential, with empirical research preliminarily validating their advantages in enhancing assessment effectiveness. Research Question: Given a range of student-related data collected through academic management systems, can we accurately predict a student's future academic performance? Method: Diverse student data are fed into the proposed model. A sparse attention mechanism is then employed to automatically identify the most discriminative features. These extracted learning features are further processed by a deep neural network, which incorporates a Cost-Sensitive Combined Loss function. A multilayer perceptron is used to derive the final grade prediction results. Findings: The experimental results show that the proposed model attains a recall rate of 74.52% across tasks. In comparison with selected baseline models, it yields recall improvements ranging from 3% to 20%, depending on the specific evaluation setting. The model is designed with considerations for deployment adaptability and data privacy, indicating its potential applicability in real-world educational scenarios.

Keywords:

Artificial intelligence

Predictive models

Feature extraction

Data models

Adaptation models

Attention mechanisms

Accuracy

Computer architecture

Transformers

Multilayer perceptrons

Cost-Sensitive Combined Loss

educational data mining

grade prediction

real-world data

sparse attention mechanism

Predicting Student Performance Using a Cost-Sensitive Deep Learning Network With Sparse Attention Mechanism

Abstract

Journal

Organization