Dots are concepts.
There are 188 fixed nodes. Concept labels are relabeled IDs, not invented skill names. The 178 visible links are a sparse projection for readability; the original native calculation uses 15,387 fixed relationships.
Scott Faust / Research, made inspectable
Inside a learner map.
One real history. A thousand interactions. A visible trail of changing evidence.
This replay preserves the original saved native map states for one previously exposed EdNet training learner. Explore all 1,000 interactions and 188 concepts, inspect the numbers, and compare direct evidence with movement introduced by neighbors.
Read the map
Structure stays. Evidence changes.
A descriptive learner-state display. The estimates summarize recorded responses; they do not establish mastery, learning gains, or an effective tutor.
Responses update evidence.
Direct estimates use cumulative correct and total counts. A response tagged to more than one concept updates each tagged concept once. All 1,001 saved positions, including the initial prior, are retained.
Neighbors can move a value.
The native 80/20 blend combines a concept’s direct estimate with a weighted mean of neighbors’ direct estimates. A smoothed value can change without direct practice. The topology does not grow or adapt during this replay.
Original real replay / minimal public projection
Follow the full thousand.
Real derived states, not synthetic data. This is an exposed training-history illustration, not a held-out predictive test or a measured learning curve. Original identifiers, items, explicit response rows, times and predictor traces are omitted. Exact state changes still retain response-pattern information; anonymity is not established.
Loading the saved 1,001-position replay…
Try two moments from the real replay:
Fixed topology · sparse visible linksAfter interaction 0 of 1,000
Direct estimate d · fixed 0–1 scale
0.000.250.500.751.00
Orange = lower; blue = higher. Dashed outline = no direct observations (n = 0); double ring = selected concept. Values and counts are numeric in the inspector and table. Equal-size circles; coordinates and links stay fixed.
Tab to a control. Enter/Space activates buttons; arrow keys step the slider one interaction; Home/End jump to the start/end. Scrubbing and changing the view pause playback.
Selected concept / complete context
Across all 1,000 interactions
The full saved trajectory is shown for context; the map at the selected position uses only the responses observed through that position. The horizontal axis is record order, not elapsed time. State changes do not establish learning gains.
━ Direct d┄ Neighbor-smoothed g│ Selected position
Inspect all 188 concepts at this position (numeric table)
| Relabeled concept | Correct s | Count n | Direct d | Smoothed g |
|---|
Under the surface
Keep the calculation and its meaning separate.
This is the original native descriptive replay, not a new simulation, fitted animation or re-run of the research pipeline.
d = (1 + s) / (2 + n)
s counts correct responses and n counts observations on a concept. The Beta(1,1) prior starts at 0.50 when n = 0; that is not evidence of 50% mastery. Exact counts and direct estimates are preserved.
g = 0.8d + 0.2 × mean
The native engine uses weighted neighboring direct estimates. It does not repeatedly smooth neighbors’ smoothed values. A changed g with unchanged n and s is indirect model movement, not a new direct observation.
Observe, then update.
Position k is after k recorded interactions and before interaction k + 1 is observed. This native path uses cumulative counts without elapsed-time forgetting. Times are omitted here; playback speed is not a learner’s practice rate.
For a smaller worked example, open the separate synthetic teaching replay. It is supplementary and does not replace this real history.
Interpretation
A changing map is a question, not a verdict.
Check
Ask what changed underneath.
Inspect counts and direct versus smoothed values. Different questions, shared tags, and neighboring evidence can all affect interpretation. A rising direct estimate can reflect easier questions, repeated items or selection—not necessarily learning.
Separate
The benchmark is a different system.
The thesis also evaluates a separate predictive representation using a 70/30 mixture, additional history features, fitted correction and calibration. That predictor is not integrated into this native map display; its traces are not included here.
Bound
Inspectability does not establish validity.
This previously exposed training example demonstrates recorded state behavior. It does not establish generalization, mastery validity, causal learning gains, teacher usefulness, improved instruction or unique graph superiority.
Source, license and adaptation
Real evidence, with an explicit scope.
Derived from EdNet KT1, published by Riiid, released under Creative Commons Attribution–NonCommercial 4.0 International for research purposes. This educational research demonstration is an adaptation by Scott Faust; no provider endorsement is implied. Dataset-derived material remains subject to that license.
The original movie selected one previously exposed training-audit learner by a fixed hash rule among histories long enough for 1,000 interactions, without ranking outcomes. This projection retains every saved native state and fixed coordinate, relabels concept IDs, and omits original identifiers, item aliases, explicit response rows, times, recent-answer arrays, forecast traces and full graph attributes. Exact counts and state changes still reveal aspects of the response pattern; this is not a guarantee of anonymity.
The original history spanned multiple sessions and gaps. This public view intentionally uses interaction order only. It is distinct from the later concept-52 gallery/thesis example; those cases and their selection rules must not be conflated.
Read the broader research overview