Programme/Deep Knowledge Tracing

⚡ Featuredprogrammev2

Deep Knowledge Tracingv2

Programme on deep knowledge tracing based on the paper by Stanford University, Khan Academy and Google

Nebular

programme creator

Skills

—

Skippable for you

ready when you are

Start Deep Knowledge Tracing

A short, four-step setup. You'll confirm enrolment + calibration, optionally run the diagnostic, wire up an agent, and land on your first session.

4 steps · ~5 min · commit on step 1

Enrolment is committed on step 1 — the rest is calibration. You can resume setup any time.

── WHAT YOU CAN NOW DO

Outcomes, not skills.

The picture of you, not of the topic. Each statement is something you'll be able to do by the end.

I can Implement and Evaluate Knowledge Tracing Models

36 skills lead here

I can Analyze and Improve Model Performance

20 skills lead here

I can Design and Justify Educational Data Strategies

21 skills lead here

I can Apply Neural Network Architectures in Knowledge Tracing

34 skills lead here

── THE CURRICULUM

3 modules. 59 skills — each with its own rubric.

The path through the material. Each lesson tackles one essential question.

your mastery model · across all skills

0 mastered0 in progress0 growth0 unexplored

L1From traces to trustworthy models

How do we turn raw student interactions into well-evaluated predictions without fooling ourselves?

L2What makes an item informative

How do item properties and skill mappings shape the probabilities we observe for student responses?

L3Learning as a hidden process

How can a simple hidden Markov view capture mastery over time and be fitted and critiqued in practice?

L4Neural sequence models for learning

Why does framing prediction as binary classification lead naturally from simple multilayer nets to gated recurrent cells?

L5How models learn what matters

How do modern sequence models learn to focus on the right context and adjust themselves through gradient signals?

L1PyTorch habits for sequence models

How do we structure code and data so a sequence model trains predictably and readably in PyTorch?

L2From interactions to recurrent state

How do clickstream events become vectors and hidden states that learn reliably over time?

L3Attention as a lens on knowledge

How can attention mechanisms both boost prediction and reveal the exercise relationships a model has discovered?

L4Why deep models for tracing

What do deep sequence models add to classic student models, and why is PyTorch a natural home for that research code?

L5Keeping evaluation honest

How do we design splits and protocols that measure generalization without leaking future information?

L6What better predictions enable

How do improved KT predictions translate to actionable teaching decisions, and where should the models evolve next?

L1Sequencing by simulated learning

How can simulated knowledge states let us compare and improve exercise orders before students ever see them?

L2Simulated students for model testing

What makes an IRT-based generator and AUC-driven validation a trustworthy stand‑in for real learners when vetting predictive models?

L3Working with the pyKT benchmark

How do pyKT’s datasets, baselines, and prediction scenarios shape what our evaluations mean in real deployments?

L4Designing fair reproducible comparisons

What makes a cross‑dataset comparison genuinely fair so that performance differences reflect models rather than accidents of setup?

L5Choosing models for real contexts

Which modeling assumptions match the realities of different learning environments?

mastered (in your learner model)in progressgrowth areanot yet exploredhover any skill for its mastery claim

── New · Unified Nebular MCP

Connect once. Every agent works.

Three ways to connect: Claude Code (PAT + install command), Claude Desktop (.mcpb download — no token to paste), or Claude web (Customise → Connectors → Add custom connector, OAuth). Same MCP endpoint, same identity on every path.

https://nebular.live/api/v1/mcp/

Generate access token Manage tokens