Data Science for Student Success: �Predicting and Supporting At-Risk Learners in the Pacific�

PRESENT BY – MR PAPPU KAPGATE – DIRECTOR OF CEIT AT SINU

CO – PRESENT BY – MR GAURAV SINGH – MASTER TRAINER OF CEIT AT SINU

The Pacific Challenge

• Hidden dropout risk emerges early but is hard to spot in time.
• Data is fragmented (paper attendance, mixed LMS, small cohorts).
• Teachers lack a simple, classroom-level “risk view.”
• Consequence: late interventions, avoidable failures, lost potential.
• Opportunity: use minimal, privacy-safe data to flag risk sooner.

Why Now

• Proven global research (OULAD) shows early signals predict outcomes.
• Lightweight tools (CSV/Sheets/LMS connectors) cut technical barriers.
• Works with small data: 30–50 students per class is enough to start.
• Privacy-by-design: pseudonymised uploads, role-based access.
• Regional momentum for practical, low-cost EdTech.

Case Study: What OULAD Proved

• Large, anonymised research dataset used to predict student outcomes.
• Data types: demographics, registration, assessments, engagement (VLE).
• Signals that worked: early assessments, engagement intensity/gaps, on-time submissions.
• Models: start simple (Logistic), boost accuracy with ensembles (Random Forest/XGBoost).
• Explanations matter: show top reasons with each risk flag.
• Takeaway: start early, mix signals, keep it explainable.

OULAD Overview

• What is OULAD?
• Public, anonymised Open University Learning Analytics Dataset used to study/predict student outcomes.
• Combines demographics, registration, assessment results, and online engagement.
• Rich longitudinal structure across modules and presentations (terms).
• Widely cited baseline for early-warning research.

OULAD Data Model (High Level)

• Six core tables & how they link
• studentInfo (who the learner is in a module/presentation)
• studentRegistration (when they join/leave)
• assessments (assessment definitions)
• studentAssessment (student scores/submissions)
• vle (virtual learning environment activities)
• studentVle (student engagement events/clicks)
• Keys: id_student, code_module, code_presentation, id_assessment, id_site.

​

What Prior Studies Found Useful

• Evidence from OULAD research�
• Early assessments are highly predictive of outcomes.
• Engagement patterns (intensity, gaps, streaks) add strong signal.
• Combining performance + engagement beats either alone.
• Logistic Regression offers transparency; Random Forest/XGBoost often raise accuracy.
• Explanations (top contributing reasons) increase trust and actionability.

​

Pacific Data Reality

• Ground truths we must design for�
• Mixed infrastructure: paper attendance, basic spreadsheets, uneven LMS usage
• Small cohorts and multi-role staff; limited IT support
• Intermittent connectivity; mobile-first reality
• Privacy sensitivity in close-knit communities
• Need for simple, repeatable processes that work offline

​

Data We Can Reliably Collect

• Practical inputs for a first-term pilot�
• Attendance: date, session, present/absent/late (paper → CSV/Sheets)
• Assessments: task name, date, weight, score, on-time/late
• Registration: enrol date, (optional) withdrawal date/reason
• Demographics (minimal): age band, prior education, disability declaration (optional)
• Engagement proxies: library/lab check-ins, activity sheets, simple LMS clicks if available

​

OULAD → Pacific Field Mapping

• Demographics:
• OULAD: gender, age_band, highest_education, disability
• Pacific proxy: same fields in ranges; keep free-text to controlled lists
• Registration:
• OULAD: date_registration, date_unregistration
• Pacific proxy: enrol date; withdrawal date (if any)
• Assessments:
• OULAD: assessment_type, date, weight; student score, date_submitted
• Pacific proxy: task type (quiz/assignment/exam), date, weight%; score; on-time/late flag
• Engagement:
• OULAD: sum_click per day/site
• Pacific proxy: attendance streaks, library/lab check-ins per day, simple LMS page views (if present)

​

Minimal Data Standard (MVP)

• Students.csv: local_student_id, age_band, highest_education, disability_flag
• Registration.csv: local_student_id, course_code, term_id, enrol_date, withdraw_date (nullable)
• Assessments.csv: course_code, term_id, assessment_id, name, type, date, weight_pct
• Scores.csv: local_student_id, assessment_id, score_pct, submitted_date, on_time_flag
• Attendance.csv: local_student_id, date, session_id, status (Present/Absent/Late)
• Engagement.csv (optional): local_student_id, date, channel (Library/Lab/LMS), count

​

Privacy & Consent Baseline

• Collect minimum necessary fields; no personal addresses/phone numbers in model tables
• Pseudonymise on upload; separate key vault from feature store
• Role-based access: classroom teachers see their classes; admins see school-level only
• Encrypted at rest & in transit; retention policy with defined purge windows
• Clear, simple consent/notice templates adapted to local norms

Risk definition & labels

• Primary label: At-risk of course non-success within the current term
• Alternatives (project selectable): Non-submission risk, Fail risk, Withdrawal risk
• Unit of prediction: Student × course × term
• Prediction horizon: Early-term (first 4–6 weeks) and mid-term checkpoints
• Output: Probability (0–1) + risk category (H/M/L)

Risk Categories for Classrooms

• High: needs action this week
• Medium: monitor & nudge
• Low: maintain momentum
• Threshold policy (MVP): tuned for higher recall with a minimum precision floor
• Each flag includes top 3 reason codes (attendance/assessment/engagement)

​

Alert Timing Windows

• Weekly batch predictions (e.g., Mondays 07:00)
• Early-term windows: Weeks 2, 3, 4 (first assessments + early attendance)
• Mid-term windows: Weeks 6, 8 (trajectory updates)
• Optional event-triggered reruns (major assessment release)
• Digest formats: class dashboard + export + email/SMS summaries (where allowed)

​

Modelling approach

• From raw data to classroom flags�
• Ingest → validate → clean → feature engineering
• Train/validation split(s) with k-fold CV
• Calibration (map scores to probabilities)
• Thresholding to H/M/L + reason codes
• Monitoring (performance, drift, coverage)

​

Model Choices

• Logistic Regression: simple, fast, interpretable baseline
• Random Forest: handles non-linearities, robust to noise
• XGBoost: strong tabular performance with tuning
• Keep baseline + champion models; compare on the same splits
• Prefer calibrated outputs for stable category thresholds

​

Feature Engineering (Pacific-ready)

• Attendance streaks (longest present streak; longest gap)
• Early assessment strength (avg of first 1–2 tasks; on-time ratio)
• Engagement trend (Δ activity week-over-week or proxy counts)
• Submission timeliness (late flags, days late)
• Cohort-relative percentiles (score/engagement vs class)
• Recency features (last 7/14 days activity, if available)

​

Explainability & Outputs

• Per-student top 3 reasons (e.g., low early score; long gap; late submission)
• Global feature importance overview for transparency
• Stable reason code taxonomy (attendance/assessment/engagement/other)
• Exportable evidence view (dates, scores, attendance summary)
• Document limits: predictions support, teachers decide

​

Platform Concept (Multi-Tenant, School-First)

• Each school/college has a private workspace (no cross-school visibility).
• Upload → validate → predict → view workflow kept simple for staff.
• Privacy by design: pseudonymise on upload, minimal fields in features.
• Modular services: ingestion, feature store, model service, dashboards.
• Admin panel for data/file history, access roles, and retention windows.

​

Secure Upload Paths

• CSV/Google Sheets/LMS connector options; guided templates for each file.
• Pre-upload checks: schema match, required columns, date formats, ID consistency.
• On ingest: pseudonymise IDs, separate key vault from model data.
• Audit trail saved for each upload (who/when/what passed/failed).
• Immediate validation report to users with fix tips; re-upload supported.

​

Processing & Storage

• Raw zone (short-lived), clean zone (validated), feature store (model-ready).
• Encryption in transit and at rest; role-based access to zones.
• Versioning: dataset & feature snapshots; reproducible model runs.
• Model registry: baseline/champion entries, metadata, and approvals.
• Automated retention: purge raw files on schedule; keep only necessary aggregates.

​

Predictions & Dashboards

• Class list view with Risk (High/Medium/Low) badges and top-3 reason codes.
• Filters: course, week, risk level; quick export (PDF/CSV).
• Student detail: brief evidence view (recent attendance, scores, engagement proxy).
• Weekly digests to staff (where permitted) with links back to dashboard.
• “Intervention notes” field for tracking actions and outcomes.

​

Offline & Low-Connectivity Modes

• Batch uploads from lab/office machines; sync when online.
• Lightweight web UI; pages sized for low bandwidth and mobile screens.
• Optional email/SMS summaries (policy-dependent) for weekly flags.
• Graceful retry/queue for failed uploads; local cache for drafts.
• Minimal footprint: run as a small VM/container or shared regional instance.

​

Teacher Workflow

• Open Class Risk view (current week).
• Sort/filter by High → Medium; scan top-3 reason codes.
• Pick actions: send message, invite to tutorial, set check-in, share study tips.
• Add intervention note (what, when, outcome).
• Revisit next week; look for risk downgrades and no-shows.
• Export class summary (PDF/CSV) for staff meetings.

​

Counsellor / Dean Workflow

• School-level dashboard: cohorts, courses, and trends across classes.
• Focus on persistently High risk students and repeat flags.
• Coordinate targeted supports: mentoring, financial/health referrals, timetable adjustments.
• Track follow-ups and close the loop with teachers.
• Monthly trend report to leadership: risk volume, actions, outcomes.

​

Governance & Access Control

• Roles, rights, and retention�
• Roles: Teacher (own classes), Counsellor/Dean (school view), Admin (config & audits).
• Least-privilege access; no cross-school visibility.
• Pseudonymised IDs in feature store; key vault isolated.
• Audit trails: uploads, predictions, threshold changes, downloads.
• Retention policy: raw uploads short-lived; features/versioned models retained per policy.
• Compliance pack: consent templates, model card, DPIA checklist.

​