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.
Presented at PiCET-2026; in press with IET Conference Proceedings · Milan Soni is the fourth of six authors.
The paper introduces the Hybrid Attention Temporal Framework (HATF) for predicting student dropout risk early in a course, from the behavioural traces a Learning Management System already collects. It combines multi-scale temporal convolution, recurrent sequence modelling and temporal attention, with Bayesian-style uncertainty estimation and fairness-aware evaluation.
This application is the product side of that work. It takes the architecture and evaluation protocol from the paper and builds the thing an institution would actually use: a REST API, an educator dashboard, an explanation for every score, and an explicit human-in-the-loop framing.
On results
Each block maps onto a component described in the paper.
weekly engagement sequence (batch, weeks, features)
│
│ multi-scale CAUSAL 1-D convolutions, kernels {1, 3, 7}
│ → very short / short / medium-term behavioural change
▼
│ unidirectional LSTM
│ → longer-range sequential dependency
▼
│ multi-head attention over weeks (padding masked)
│ → which weeks this assessment rests on
▼
context vector (+ residual last state, + week & course embeddings)
│
│ MLP head, dropout kept ACTIVE at inference
│ → Monte Carlo Dropout posterior over the risk score
▼
risk probability + uncertainty + attention mapResearch
Architecture, early-prediction protocol, fairness-aware evaluation.
Engineering
Reproducible synthetic data, leakage-controlled feature pipeline, one checkpoint serving every prediction week, calibrated probabilities, REST API, dashboard.
Product
An explanation for every score, uncertainty surfaced instead of hidden, human-in-the-loop framing, supportive language, concrete advisor actions.
Measured on held-out test students. Synthetic data.
ROC-AUC
0.845
Pooled across every prediction week.
F1
0.693
Calibration error
0.038
Expected Calibration Error after Platt scaling.
Worst parity gap
0.200
Largest in academic_background.
Decision support for educators and academic advisors: surface students who may benefit from timely, proactive outreach.
Responsible AI
Causal convolutions
Sensitive attributes
course_id. Fairness is then measured across gender, age group and academic background rather than assumed away.Out of scope
Smoothed F2-optimal on the validation split, subject to flagging at most 40% of the cohort. F2 weights recall twice as heavily as precision because missing a student who needed support is a worse error than an unnecessary check-in; the flag-rate cap keeps the result actionable.