Automatic recognition and assessment of students’ high-order thinking skills in online learning environments based on multimodal deep learning

Scalable prediction of rubric-consistent proxy indicators related to students’ high-order thinking skills in online learning environments is important for supporting process-oriented learning analytics and personalized educational intervention. However, existing assessment methods often rely on manual scoring, single-source learning evidence, or shallow multimodal fusion strategies, which makes it difficult to capture the complex relationships among students’ textual expression, learning behavior, task performance, and interaction patterns. To address these limitations, this paper proposes MHTA-Net, a multimodal network for predicting rubric-derived high-order thinking proxy targets through multimodal encoding, Mamba-based selective fusion, and KAN-based multi-task prediction. Specifically, the multimodal learning feature encoder extracts heterogeneous representations from textual expressions, behavioral sequences, assessment records, and interaction features. The Mamba-based selective fusion module models cross-modal dependencies and selectively retains informative evidence states from heterogeneous learning traces, while the KAN-based multi-task module maps the fused representation to an overall rubric-derived proxy level and dimension-specific proxy scores for analysis, evaluation, and creation. Experiments are conducted on the constructed MOOCCubeX-HOTS benchmark, which contains 12,486 learner-course instances derived from public MOOC data. The proposed MHTA-Net achieves the best overall performance, with an Accuracy of 0.876, a Macro-F1 of 0.862, an AUC of 0.923, an MAE of 4.96, an RMSE of 6.42, and a PCC of 0.842. Ablation experiments, dimension-level analysis, missing-modality robustness evaluation, visualization, and case studies further demonstrate the predictive effectiveness, robustness, and output transparency of the proposed model. The results indicate that MHTA-Net provides a feasible framework for multidimensional prediction of rubric-derived proxy indicators related to high-order thinking in online learning environments, rather than an independently validated measurement of latent cognitive ability.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-18
DOI
https://doi.org/10.1038/s41598-026-79580-1
Primary Topic
Online Learning and Analytics
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Automatic recognition and assessment of students’ high-order thinking skills in online learning environments based on multimodal deep learning

Zongyi Liu, Hui Li, Guo Li Zhou
1 links
Scientific Reports
Online Learning and Analytics
article

Automatic recognition and assessment of students’ high-order thinking skills in online learning environments based on multimodal deep learning

Zongyi Liu, Hui Li, Guo Li Zhou
article en

Abstract

Scalable prediction of rubric-consistent proxy indicators related to students’ high-order thinking skills in online learning environments is important for supporting process-oriented learning analytics and personalized educational intervention. However, existing assessment methods often rely on manual scoring, single-source learning evidence, or shallow multimodal fusion strategies, which makes it difficult to capture the complex relationships among students’ textual expression, learning behavior, task performance, and interaction patterns. To address these limitations, this paper proposes MHTA-Net, a multimodal network for predicting rubric-derived high-order thinking proxy targets through multimodal encoding, Mamba-based selective fusion, and KAN-based multi-task prediction. Specifically, the multimodal learning feature encoder extracts heterogeneous representations from textual expressions, behavioral sequences, assessment records, and interaction features. The Mamba-based selective fusion module models cross-modal dependencies and selectively retains informative evidence states from heterogeneous learning traces, while the KAN-based multi-task module maps the fused representation to an overall rubric-derived proxy level and dimension-specific proxy scores for analysis, evaluation, and creation. Experiments are conducted on the constructed MOOCCubeX-HOTS benchmark, which contains 12,486 learner-course instances derived from public MOOC data. The proposed MHTA-Net achieves the best overall performance, with an Accuracy of 0.876, a Macro-F1 of 0.862, an AUC of 0.923, an MAE of 4.96, an RMSE of 6.42, and a PCC of 0.842. Ablation experiments, dimension-level analysis, missing-modality robustness evaluation, visualization, and case studies further demonstrate the predictive effectiveness, robustness, and output transparency of the proposed model. The results indicate that MHTA-Net provides a feasible framework for multidimensional prediction of rubric-derived proxy indicators related to high-order thinking in online learning environments, rather than an independently validated measurement of latent cognitive ability.

Scientific Reports
National University of Malaysia (MY)
Quality education
Online Learning and Analytics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon