A practical onboarding guide for mapping educational technology and learning-sciences data to defensible analyses, claim verbs, research genres, and reviewer checks before the first model is run.
Educational data are produced by platforms, tasks, learners, teachers, sensors, institutions, and researchers. The same dataset can support a descriptive claim, a process claim, a measurement claim, or a prediction claim, but not all at once and not without different evidence burdens.
Name the data type and decide what kind of claim it can support.
Map data sources to method families and reviewer-safe claim language.
A defensible analysis plan, not just a list of popular techniques.
These videos fit the guide's central topic: how educational data become learning analytics, educational data mining, and learning-sciences evidence. Use them as conceptual anchors before moving into the matrix.
A SoLAR primer that frames learning analytics as data about learners and learning, interpreted for teaching and learning action. Best before using the matrix because it clarifies the field-level purpose.
Open on YouTubeA practical EDM overview focused on how logs and platform data can be analyzed to understand learning behavior, prediction, and improvement. Pair it with the trace, sequence, and prediction rows.
Open on YouTubeA SoLAR LASI panel connecting analytics with learning-sciences questions. This is the best fit for remembering that data analysis must preserve theory, context, and learning mechanisms.
Open on YouTubeThe selector gives a first-pass diagnosis. It does not approve a study; it tells you what to inspect before writing the method section.
Summaries, dashboards, distributions, heatmaps, and temporal plots. Best for pattern visibility.
Supervised models with leakage-free validation, calibration, and subgroup checks.
Reliability, IRT/Rasch, cognitive diagnosis, validity arguments, and evidence-centered design.
Sequence mining, TNA, ONA, process mining, Markov models, and HMMs.
Qualitative coding, NLP, embeddings, discourse moves, argumentation, and LLM-assisted coding with validation.
Social networks, two-mode networks, epistemic networks, centrality, community detection, and group structure.
Synchronization, feature extraction, early/late fusion, alignment, and triangulation across streams.
RCT, QED, RD, ITS, SCED, DiD, and causal models only when the design carries the claim.
Joint displays, explanatory sequential, convergent integration, and meta-inferences.
These are lightweight visualization patterns that can turn a methods guide into a more playful research-planning tool. Each pattern keeps the same discipline: the visual must reveal a data-to-method-to-claim relationship, not decorate it.
Design rule: readers should be able to click a data source and immediately see which methods and claims become safer or riskier.
Do not treat clicks, dwell time, facial expression, or log volume as direct engagement without a construct argument and triangulation or validation evidence.
Show held-out validation or honest cross-validation, outcome timing, leakage checks, calibration, and subgroup performance.
Name the design feature that supports causal interpretation: randomization, quasi-experimental equivalence, regression discontinuity, interrupted time series, single-case design, or another credible strategy.
Counts of posts, replies, or centrality do not automatically indicate collaboration quality. Tie definitions and discourse meaning must be visible.
Multiple data streams do not automatically triangulate. State exactly where logs, text, video, sensors, surveys, or outcomes inform one another.
The fastest workflow is not to ask for a finished webpage in one sentence. Treat Claude Code or Codex as a research assistant, editor, designer, and publisher, but keep the evidence rules explicit at every step.
Ask the agent to define the audience, data families, claim verbs, and reviewer risks before writing UI code.
Have it collect primary references and remove any citation it cannot verify.
Request an English GitHub Pages page with interactive selectors, visual ideas, downloads, and thumbnail metadata.
Make the agent test JavaScript, inspect the rendered DOM, commit, push, and report the public URL.