Decision support, not a verdict. Risk flags indicate that a student may benefit from timely support and should be reviewed by an educator or advisor. All data shown is synthetic.
Department of Computer Science & Engineering, Global Institute of Technology, Jaipur, Rajasthan, India
Listed in publication order. Department of Computer Science & Engineering, Global Institute of Technology.
Pradeep Jha
Author 1 of 6
Manju Mathur
Author 2 of 6
Abhay Purohit
Author 3 of 6
Milan Soni
Author 4 of 6
Avadhi Singhal
Author 5 of 6
Presented the paper at the conference
Abhilash Joshi
Author 6 of 6
8th Parul University International Conference on Engineering & Technology
PiCET-2026
“Innovations in Computing: Smart, Sustainable and Emerging Technologies”
Faculty of Engineering & Technology, Parul University
Publication details
The abstract, in plain language.
Online and blended courses lose a large share of their students, and the warning signs move week to week — but most prediction models treat engagement as static and miss those changes.
The paper proposes the Hybrid Attention Temporal Framework (HATF): multi-scale temporal modelling to catch both sudden and slow behavioural change, attention to identify which periods mattered, adaptive time windows instead of one fixed window for every course, Bayesian uncertainty estimation so predictions come with a confidence measure, and fairness-aware training across student groups.
It was evaluated on three real LMS datasets. The attention maps also gave educators a readable account of which learning periods drive dropout risk.
Results on three real LMS datasets. These are the paper's numbers, not this project's — the two are measured on completely different data and are not comparable.
Computer science
2,847
students across 15 courses
Business administration
1,932
students across 12 courses
Engineering
3,156
students across 18 courses
HATF against the baselines (paper, Table 2)
All figures are percentages, on 7,935 real students across 45 courses.
| Model | Accuracy | Precision | F1 | AUC |
|---|---|---|---|---|
| Logistic Regression | 78.3 | 76.1 | 77.6 | 82.4 |
| Random Forest | 82.7 | 80.9 | 82.3 | 87.1 |
| LSTM | 85.9 | 84.2 | 85.4 | 89.8 |
| Attention–CNN | 87.3 | 86.1 | 87.1 | 91.2 |
| FWTS–CNN | 89.1 | 87.8 | 88.8 | 92.5 |
| HATF (proposed) | 92.4 | 91.2 | 94.2 | 96.1 |
Early prediction (paper, Figure 2)
The point of the framework — usable results before the course ends.
Fairness (paper, Table 3)
Parity difference is best near 0; the odds ratio is best near 1.
The paper also reports a calibration error of 0.034, with 87% accuracy on high-confidence predictions falling to 73% on low-confidence ones — which is the behaviour you want from an uncertainty estimate.
A note on one inconsistency
Component by component, with an honest label on each: implemented as described, one possible reading of an underspecified idea, or only partly done.
Multi-scale temporal encoder
Implemented as describedIn the paper
Conv1D with kernel sizes {1, 3, 7}, feeding an LSTM
In this codebase
Implemented exactly, with causal (left-only) padding added so a truncated week can never see the future
Adaptive time windows
One reading of itIn the paper
Flexible, information-driven windows instead of one fixed size
In this codebase
A learned softmax gate over the kernel scales, computed per student and per week, surfaced in the UI
Multi-head attention
Implemented as describedIn the paper
8 heads over the LSTM output; context vector = attention × output
In this codebase
Implemented exactly; the prediction head reads the context vector directly, as in the paper
Bayesian prediction layer
Implemented as describedIn the paper
Monte Carlo Dropout, 100 samples, for uncertainty
In this codebase
100 samples on the full profile (30 on the fast demo profile), plus Platt calibration
Feature families
Implemented as describedIn the paper
Activity, Assessment, Interaction, Temporal
In this codebase
All four, 25 engineered features
Fairness-aware training
Partly implementedIn the paper
Demographic parity, equalised odds, individual fairness as loss constraints
In this codebase
All three measured; demographic parity added as an in-batch training penalty
Early prediction from week 2
Implemented as describedIn the paper
P(dropout | X₁:t), minimum t = 2
In this codebase
One model serves every week from 2 onward, trained on truncated sequences
Baseline comparison
Partly implementedIn the paper
LR, RF, SVM, LSTM, CNN, CNN-LSTM, Transformer, Attention-CNN, FWTS-CNN, Time-Aware Rules
In this codebase
8 of the 10 reproduced under identical conditions; FWTS-CNN and Time-Aware Rules omitted as they need their source papers' specifics
Why this project's numbers are lower than the paper's
The paper measures HATF on 7,935 real students. This project runs on 500 invented students, because using real student records in a portfolio project would be a privacy problem, not an achievement.
Simulated behaviour is simpler than real behaviour, so there is less for a deep model to find — which is why the benchmark here shows HATF losing to simpler models, and why that result is published rather than hidden. It demonstrates that the system works end to end; it cannot and does not verify the paper's findings.
@inproceedings{jha2026hatf,
title = {Hybrid Attention-Based Temporal Modeling for Early Dropout Prediction in Learning Management Systems},
author = {Pradeep Jha and Manju Mathur and Abhay Purohit and Milan Soni and Avadhi Singhal and Abhilash Joshi},
booktitle = {8th Parul University International Conference on Engineering & Technology (PiCET-2026)},
address = {Jaipur, Rajasthan, India},
year = {2026},
month = {May},
publisher = {IET Conference Proceedings},
note = {Scopus indexed. Paper ID PU/PiCET26/COP/327}
}