Temporal deep learning and multitask learning offer no advantage over tree ensembles for early warning of student course outcomes

Abstract We test whether multi-task learning improves prediction of pass/fail and withdrawal from VLE logs—it does not. Using 14-day interaction sequences from 26,717 Open University students, we compare single-task and multi-task BiLSTM architectures against gradient boosting baselines across 10 runs. All deep learning variants achieved statistically equivalent F1 scores: 0.600 ± 0.004 (MTL without cross-attention), 0.598 ± 0.003 (MTL with cross-attention), and 0.597 ± 0.006 (single-task), with pairwise 95% CIs falling within [-0.008, 0.003] (TOST p < 0.001, δ = 0.02). BiLSTM outperformed LightGBM by 2.2% (t = 7.26,13 p < 0.001, d = 0.73). Despite moderate task correlation (r = 0.61), MTL provided no advantage over single-task training—a finding robust to excluding withdrawn students (F1 change: -0.6%). In this dataset, task relatedness alone (r = 0.61) did not ensure multi-task benefits: pass/fail and withdrawal appear to have partially distinct behavioral antecedents, limiting cross-task transfer.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-14
DOI
https://doi.org/10.1038/s41598-026-80425-0
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

Temporal deep learning and multitask learning offer no advantage over tree ensembles for early warning of student course outcomes

Zhang Chuanxia, Linkai qi, Haili Qin
2 links
Scientific Reports
Online Learning and Analytics
article

Temporal deep learning and multitask learning offer no advantage over tree ensembles for early warning of student course outcomes

Zhang Chuanxia, Linkai qi, Haili Qin
article en

Abstract

Abstract We test whether multi-task learning improves prediction of pass/fail and withdrawal from VLE logs—it does not. Using 14-day interaction sequences from 26,717 Open University students, we compare single-task and multi-task BiLSTM architectures against gradient boosting baselines across 10 runs. All deep learning variants achieved statistically equivalent F1 scores: 0.600 ± 0.004 (MTL without cross-attention), 0.598 ± 0.003 (MTL with cross-attention), and 0.597 ± 0.006 (single-task), with pairwise 95% CIs falling within [-0.008, 0.003] (TOST p < 0.001, δ = 0.02). BiLSTM outperformed LightGBM by 2.2% (t = 7.26,13 p < 0.001, d = 0.73). Despite moderate task correlation (r = 0.61), MTL provided no advantage over single-task training—a finding robust to excluding withdrawn students (F1 change: -0.6%). In this dataset, task relatedness alone (r = 0.61) did not ensure multi-task benefits: pass/fail and withdrawal appear to have partially distinct behavioral antecedents, limiting cross-task transfer.

Scientific Reports
Shanghai University of Engineering Science (CN), Shandong Xiehe University (CN)
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