Research Project  ·  SLIIT  ·  2025–2026

Adaptive & Explainable AI for
|

AcademiGuard is a proactive student monitoring platform that detects early signs of disengagement before academic performance suffers — powered by a GRU Autoencoder, Reinforcement Learning, and SHAP-based Explainable AI for full educator transparency.

SLIIT · Dept. of Information Technology Supervisor: Sanjeevi Chandrasiri Co-Supervisor: Ishara Weerathunga
Model Performance
0%
Ensemble Accuracy
0%
ROC-AUC Score
0K+
LMS Records
0%
Cross-Val Mean
TECHNOLOGY STACK
React Firebase FastAPI GRU SHAP/XAI XGBoost LightGBM LLaMA
Scroll to explore

Three Pillars of AcademiGuard

An integrated architecture unifying predictive modeling, explainable AI, and adaptive cognitive support.

Hybrid Ensemble Engine

Random Forest + XGBoost + LightGBM (2:2:1 soft-voting) achieves 97.8% accuracy identifying at-risk students from 17-attribute academic and demographic data.

97.8%Accuracy

GRU Anomaly Detector

A GRU Autoencoder monitors 10-week LMS behavioral sequences and flags students exceeding the 97th percentile reconstruction error as high-risk for disengagement.

p97Risk Threshold
🎯

RL Intervention Agent

A Q-learning agent prescribes optimal re-engagement — soft nudges (40.9%), reminders (6.8%), or escalation — while LLaMA delivers personalized, empathetic messaging.

40.9%Soft Nudge Rate

Research Domain

Bridging Educational Data Mining, Explainable AI, and Reinforcement Learning to build a trustworthy end-to-end academic early-warning ecosystem.

📚

Literature Survey

Educational Data Mining (EDM) and Learning Analytics (LA) have emerged as critical research domains [1][2][19]. With widespread LMS adoption, behavioral footprints — login frequencies, assessment results, assignment patterns, and session durations — are now accessible at scale and proven to be highly reliable early indicators of academic trajectory [3].

Supervised ML techniques (decision trees, SVMs, random forests, deep neural networks) achieve strong predictive accuracy [3][15][16]. Temporal models such as GRU and LSTM networks further capture sequential behavioral patterns for more accurate at-risk identification [4][5], while autoencoder architectures enable unsupervised anomaly detection in LMS activity streams [6]. However, their "black-box" nature limits classroom adoption — educators are hesitant to act on risk warnings they cannot interpret. SHAP (SHapley Additive exPlanations) has emerged as the leading approach to bridge this gap [8][10][11]. Learner disengagement patterns in online environments have been studied extensively, particularly in MOOC contexts [7][9][20].

Recent work also demonstrates that reinforcement learning agents can prescribe adaptive, personalized interventions by optimizing long-term re-engagement outcomes [12][13][14], yet no unified platform combines real-time anomaly detection, explainability, and adaptive interventions within a scalable cloud architecture [17][18].

References
  1. M. Yağcı, "Educational data mining: Prediction of students' academic performance using machine learning algorithms," Smart Learning Environments, vol. 9, no. 1, pp. 1–19, 2022.
  2. G. Siemens and R. S. Baker, "Learning analytics and educational data mining: Towards communication and collaboration," in Proc. Learning Analytics and Knowledge Conf. (LAK), 2012.
  3. A. M. Shahiri, W. Husain, and N. A. Rashid, "A review on predicting student's performance using data mining techniques," Procedia Computer Science, vol. 72, pp. 414–422, 2015.
  4. X. Chen, Y. Liu, and M. Zhao, "A novel student performance prediction model based on GRU and attention mechanism," IEEE Access, vol. 11, pp. 23450–23461, 2023.
  5. R. Mubarak, H. Wang, and S. M. Zin, "Predicting student performance using LSTM and GRU sequence models," IEEE Transactions on Learning Technologies, vol. 15, no. 2, pp. 190–202, 2022.
  6. L. Zhang et al., "Anomaly detection in student learning behaviors using autoencoder architectures," IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3412–3424, 2021.
  7. S. Kizilcec, C. Piech, and E. Schneider, "Deconstructing disengagement: Analyzing learner behavior in MOOCs," in Proc. Learning Analytics and Knowledge Conf. (LAK), 2013.
  8. S. H. Lundberg and S. Lee, "A unified approach to interpreting model predictions," in Advances in Neural Information Processing Systems (NeurIPS), 2017.
  9. B. Shneiderman, "Human-centered artificial intelligence," Communications of the ACM, vol. 63, no. 1, pp. 58–61, 2020.
  10. J. Wang, T. Li, and C. K. Looi, "Explainable AI in Educational Data Mining: Using SHAP to interpret predictive models," IEEE Transactions on Education, vol. 67, no. 1, pp. 45–56, 2024.
  11. M. A. Alam and M. S. Hossain, "Trustworthy learning analytics: Integrating feature attribution methods for student retention," IEEE Access, vol. 10, pp. 115023–115035, 2022.
  12. Y. Zheng et al., "Reinforcement Learning for Personalized Education: Adaptive Intervention Strategies," IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 8, pp. 8120–8134, 2023.
  13. A. N. Rafferty, H. Ying, and P. A. Williams, "Using reinforcement learning to design personalized educational interventions," Journal of Educational Data Mining, vol. 13, no. 3, 2021.
  14. D. Zhou and K. Koedinger, "Dynamic decision making in intelligent tutoring systems via Markov Decision Processes," IEEE Transactions on Learning Technologies, vol. 14, no. 4, pp. 512–525, 2021.
  15. T. Chen and C. Guestrin, "XGBoost: A scalable tree boosting system," in Proc. 22nd ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining (KDD), pp. 785–794, 2016.
  16. G. Ke et al., "LightGBM: A highly efficient gradient boosting decision tree," in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, pp. 3146–3154, 2017.
  17. A. Al-Shabandar, A. Jaddoa, P. Liatsis, and A. J. Hussain, "A cloud-based educational analytics architecture using serverless backend services," IEEE Transactions on Cloud Computing, vol. 10, no. 2, pp. 1021–1034, 2022.
  18. S. V. Kumar and R. P. Singh, "Scalable Microservice Architectures for Real-Time Educational Dashboards: Integrating Fast APIs and Reactive Web Frameworks," in IEEE/ACM International Conference on Automated Software Engineering (ASE), 2023.
  19. C. Romero and S. Ventura, "Educational data mining: A review of the state of the art," IEEE Transactions on Systems, Man, and Cybernetics, Part C, vol. 40, no. 6, pp. 601–618, 2010.
  20. K. Verbert, E. Duval, J. Klerkx, S. Govaerts, and J. L. Santos, "Learning dashboards: An overview and future research opportunities," Personal and Ubiquitous Computing, vol. 18, no. 6, pp. 1499–1514, 2014.
🔍

Research Gap

Despite significant progress across individual research domains, a critical gap persists: the majority of existing systems remain siloed. They either focus exclusively on risk prediction without explainability, or lack the infrastructure to trigger adaptive personalized interventions automatically.

Three specific gaps were identified during our literature review:

  • Risk prediction systems lack integrated explainability — educators receive alerts but not the reasoning behind them
  • Anomaly detection frameworks operate in isolation, without downstream adaptive intervention logic
  • Intervention systems are static and rule-based, unable to learn from individual student response patterns over time

AcademiGuard directly addresses this gap by synthesizing sequential disengagement detection, SHAP-driven explainability, and adaptive intervention recommendations into a single cohesive platform.

⚠️

Research Problem

Traditional academic monitoring relies on lagging indicators — primarily midterm and final grades. Interventions based on these metrics occur too late to meaningfully alter a student's academic trajectory.

A student who gradually reduces forum participation, exhibits irregular login patterns, or consistently submits assignments late is often already on a dropout trajectory — long before grades reflect the problem. Universities frequently fail to catch these early warning signs.

Core Research Question: How can we detect and respond to student disengagement proactively — using real-time behavioral signals — in a way that is both effective and transparent to educators?

This requires bridging early behavioral anomaly detection, prescription of optimal intervention strategies, and human-interpretable reasoning behind every automated recommendation.

🎯

Research Objectives

01
Scalable Architecture

Deploy a full-stack cloud educational monitoring system integrating React, Firebase, and Python microservices via FastAPI.

02
Ensemble Risk Prediction

Build a Hybrid Ensemble (RF + XGBoost + LightGBM, 2:2:1 soft-voting) for baseline cumulative academic risk assessment.

03
Temporal Anomaly Detection

Apply a GRU Autoencoder to detect sequential behavioral anomalies in LMS data before grades reflect academic decline.

04
Adaptive Interventions

Use a Q-learning RL agent to prescribe optimal, escalating re-engagement strategies based on student risk profiles and response history.

05
Explainable AI

Embed SHAP-based XAI to provide per-feature attribution scores, ensuring every risk prediction is interpretable and actionable for educators.

⚙️

Methodology

AcademiGuard operates as a dual-engine ML framework with three integrated stages:

1
Hybrid Ensemble Engine (Baseline Risk)

Soft-voting classifier aggregates Random Forest (300 estimators, depth 12), XGBoost (lr=0.05, scale_pos_weight=4), and LightGBM (2:2:1 weights). Trained on 17-attribute UCI dataset with 80:20 stratified split and 5-fold cross-validation. SHAP TreeExplainer surfaces per-feature contributions for every prediction.

2
GRU Autoencoder (Continuous Monitoring)

Trained on 75,000 weekly LMS records across 5,000 students. Compresses 10-week behavioral sequences via Adam optimizer (MSE loss). Normal patterns are learned; reconstruction error spikes identify anomalous, disengaging students. Students exceeding the 97th percentile threshold trigger the intervention module.

3
Q-Learning RL Agent (Intervention)

Trained on 25,000 historical interaction samples. State space: risk_level, risk_trend, last_action, no_response_streak, fatigue_level. Selects from DO_NOTHING (52.3%), SOFT_NUDGE (40.9%), and REMINDER (6.8%). LLaMA translates decisions into personalized empathetic natural-language messages.

🛠️

Technologies Used

Frontend

React.jsTailwind CSS

Backend

Firebase (Auth / Firestore)FastAPI (Python)RESTful API Gateway

Machine Learning

GRU AutoencoderRandom ForestXGBoostLightGBMQ-Learning (RL)MinMaxScalerAdam Optimizer

Explainability

SHAP TreeExplainerFeature Attribution Scores

AI / NLP

LLaMA (Cognitive AI)Natural Language Generation

Data Sources

UCI ML RepositoryUniversity LMS LogsFirestore Database

Project Timeline

Key assessments and deliverables throughout the research lifecycle. Use the dropdown below to navigate to a specific milestone.

May
2025
Project Topic AssessmentNo Marks

Initial topic assessment submitted with a brief overview of the proposed research, problem statement, objectives, and member task breakdown.

✓ Completed  ·  30 May 2025
Jun
2025
TAF SubmissionNo Marks

Team Agreement Form submitted outlining scope, objectives, overall solution architecture, and member responsibilities for AcademiGuard.

✓ Completed  ·  24 Jun 2025
Jul
2025
Project Charter Form SubmissionNo Marks

Submission of the Project Charter Form formally defining the project scope, objectives, team roles, and initial planning framework for AcademiGuard.

✓ Completed  ·  23 Jul 2025
Sep
2025
Project Proposal Presentation6%

Presented to a panel of judges covering AcademiGuard's proposed dual-engine architecture, methodology, and expected research contributions.

✓ Completed  ·  11 Sep 2025
Sep
2025
Project Proposal Report6%

In-depth analysis report covering the research problem, literature review, proposed solution architecture, methodology, and expected contributions to EDM and LA.

✓ Completed  ·  19 Sep 2025
Jan
2026
Check List 1 Submission1%

Document overviewing key Phase 1 implementation tasks completed by each team member, including the Hybrid Ensemble engine and initial dataset preprocessing.

✓ Completed  ·  05 Jan 2026
Jan
2026
Progress Presentation 115%

Evaluation of 50% project completion demonstrating the Hybrid Ensemble engine achieving 97.8% accuracy and the early GRU prototype to the assessment panel.

✓ Completed  ·  06 Jan 2026
Mar
2026
Progress Presentation 218%

Evaluation of 90% project completion with full platform demonstration: GRU anomaly detection, RL interventions, SHAP visualizations, and LLaMA-powered messaging.

✓ Completed  ·  09 Mar 2026
Mar
2026
Check List 2 Submission1%

Document overviewing key Phase 2 implementation tasks completed by each team member, including the GRU Autoencoder, RL intervention agent, and SHAP explainability layer.

✓ Completed  ·  09 Mar 2026
Apr
2026
Draft Thesis SubmissionNo Marks

Submission of the draft thesis covering the complete AcademiGuard research, implementation details, experimental evaluation, and preliminary conclusions.

✓ Completed  ·  26 Apr 2026
Apr
2026
Website Submission2%

Submission of the AcademiGuard research project website showcasing all project components, team information, documents, milestones, and research outcomes.

✓ Completed  ·  26 Apr 2026
May
2026
Final Presentation & Viva20%

Final project evaluation and viva examination demonstrating the complete AcademiGuard platform and defending research contributions before the assessment panel.

⏳ Pending  ·  05 May 2026
May
2026
Website Evaluation & Logbook Submission2%

Submission of the evaluated research website alongside the project logbook documenting the development and research progress throughout the year.

⏳ Pending  ·  05 May 2026
May
2026
Research Paper Submission10%

Formal submission of the IEEE-format research paper documenting the complete AcademiGuard platform, methodology, and experimental results for publication review.

⏳ Pending  ·  08 May 2026
May
2026
Final Thesis Submission19%

Submission of the final individual and group thesis documents covering the complete AcademiGuard research, implementation, evaluation, and conclusions.

⏳ Pending  ·  13 May 2026
Jun
2026
Research Paper Publication Evidence SubmissionNo Marks

Submission of evidence confirming the research paper has been accepted or published in an IEEE conference or journal proceeding.

⏳ Pending  ·  15 Jun 2026

Project Documents

All project documents grouped by category. Click a row to expand, then view or download individual files.

📋
TAF Submission
Team Agreement Form  ·  Jun 2025  ·  No Marks
1 file
PDF
TAF_25-26J-172.pdf
Group  ·  Team Agreement Form
📄
Research Proposals
Individual Proposals  ·  Sep 2025  ·  6%
4 files
PDF
Ravisanka U.V.P  ·  IT22354792
Individual  ·  Research Proposal
PDF
Disanayaka S.T  ·  IT22370228
Individual  ·  Research Proposal
PDF
Nimanji D.L.K  ·  IT22365750
Individual  ·  Research Proposal
PDF
Perera I.A.T.D  ·  IT22902702
Individual  ·  Research Proposal
Check Lists
Progress Checklists  ·  Jan & Mar 2026  ·  2% total
2 files
TXT
Check List 1
Group  ·  Jan 2026  ·  1%
XLSX
Check List 2
Group  ·  Mar 2026  ·  1%
📰
IEEE Research Paper
Conference Paper  ·  Mar 2026  ·  10%
1 file
PDF
IEEE Research Paper.pdf
Group  ·  AcademiGuard Full Paper
📚
Final Reports
Group & Individual Reports  ·  Apr 2026  ·  19% total
5 files
PDF
Final Report  ·  Group
Group  ·  4%  ·  All Members
PDF
Ravisanka U.V.P  ·  IT22354792
Individual  ·  15%
PDF
Disanayaka S.T  ·  IT22370228
Individual  ·  15%
PDF
Nimanji D.L.K  ·  IT22365750
Individual  ·  15%
PDF
Perera I.A.T.D  ·  IT22902702
Individual  ·  15%
📗
Logbook
Project Logbook  ·  May 2026  ·  2%
Pending Due 05 May
🎉
Research Paper Publication Evidence
Acceptance / Publication Proof  ·  Jun 2026  ·  No Marks
Pending Due 15 Jun

Slides & Presentations

Presentation slides from past assessments and future milestones.

Meet the Team

Information Technology undergraduates at SLIIT, dedicated to making AI-driven academic support transparent, proactive, and trustworthy.

SC
Supervisor
Sanjeevi Chandrasiri
Dept. of Information Technology, SLIIT
LinkedIn
IW
Co-Supervisor
Ishara Weerathunga
Dept. of Information Technology, SLIIT
LinkedIn
Research Group Members
RU
Team Leader
Ravisanka U.V.P
IT22354792
LinkedIn
DS
Team Member
Disanayaka S.T
IT22370228
LinkedIn
ND
Team Member
Nimanji D.L.K
IT22365750
LinkedIn
PI
Team Member
Perera I.A.T.D
IT22902702
LinkedIn

Get in Touch

Have questions about AcademiGuard? Reach out to our team or supervisors for collaboration or further information.

🏫

Institution

Sri Lanka Institute of Information Technology (SLIIT), Malabe, Sri Lanka

📧

Supervisor

sanji.c@sliit.lk

📧

Co-Supervisor

ishara.w@sliit.lk

🏛️

Department

Department of Information Technology

Send a Message