EduRAG：用因果AI解释学生表现背后的真正驱动因素

EduRAG: Using Causal AI to Explain Real Drivers Behind Student Performance

Abstract: A prototype that transforms LMS behavior data into causal-driven, actionable intervention suggestions instead of black-box risk scores.

Source Info:

• Original Author/Maintainer: TreblaMagic
• Source Platform: GitHub
• Original Title: EduRAG — Causal AI for Student Success
• Link: https://github.com/TreblaMagic/EduRAG
• Updated: 2026-05-30T13:45:55Z

Core Goal: Bridge the gap between predictive risk scoring and causal insight in learning analytics, enabling educators to understand why students are at risk and what actions to take.

Most existing learning analytics tools only answer 'who will fail' via risk scores but cannot explain 'why' or 'what to do'. This black-box approach limits educators' ability to make effective interventions.

EduRAG was created to fill this gap: it is not just a prediction system but an interpretable, intervention-oriented platform that turns LMS data into causal-driven action suggestions.

EduRAG's core innovation is its double-layer design:

1. Prediction Layer: Answers 'who is at risk?' using logistic regression on LMS features (login frequency, assignment submission time, video watch duration) to compute risk scores.

2. Causal Layer: Answers 'how will adjusting a factor change outcomes?' using Structural Causal Modeling (SCM), Directed Acyclic Graph (DAG) reasoning, backdoor adjustment (OLS regression), bootstrap confidence intervals, and refutation checks to identify actionable factors affecting student performance.

Tech Stack:

• Frontend: Next.js15 (App Router) + React19 + Tailwind CSS
• Backend: Next.js server actions/routes + Prisma ORM
• Database: SQLite (MVP) → compatible with PostgreSQL
• Causal Engine: TypeScript (custom OLS, bootstrap sampling, PC algorithm) + optional Python (DoWhy, causal-learn)
• Prediction Engine: TypeScript L2 logistic regression + optional Python (scikit-learn LR/RF)
• Visualization: Custom SVG charts/DAG renderer (no external libraries)

Data Pipeline: Supports synthetic CSV (generated via npm run data:generate), Shell University simulation (via npm run shell:seed), and real CSV upload (with validation/preview/submit flow).

Core Features:

• Student Profile: Timeline, prediction+intervention panels, counterfactual intervention cards (sorted by impact)
• Causal Graph Compare: Side-by-side view of manual (expert) vs algorithm-discovered (PC) DAGs
• What-If Simulator: Adjust variables (e.g., video time, assignment submission time) to see outcome changes with bootstrap CIs
• Intervention Feedback Loop: Track decisions (accept/reject/postpone/complete) and outcomes

EduRAG redefines educational analytics:

• From 'Who' to 'Why': Instead of just '70% fail risk', it explains: 'Low video watch time and late first assignment cause the risk. Increasing video time by20% raises pass rate to85%.'

• From Prediction to Intervention: Clearly distinguishes prediction (identifying risk) from intervention (solving it), critical for improving educational equity and effectiveness.

• Interpretability: Essential in education—educators need to understand why a suggestion is made to trust and act on it.

Current Phase: 12A (GitHub Ready). Key milestones:

• Phase5-9: Zero new runtime dependencies; hand-coded features (charts, DAG renderer, causal estimator, PC algorithm)
• Phase10: Dataset switcher (synthetic/Shell University/upload)
• Phase11: Intervention feedback loop
• Phase12A: GitHub ready
• Phase12D: First deployment上线

EduRAG represents a key shift in edtech from correlation to causality. As a prototype, it provides a valuable reference for the field.

Future: More causal AI systems will help educators not just 'see' problems but 'understand' and 'solve' them—unlocking AI's true value in empowering human decision-making.