Skip to content

Prediction Fairness Explorer

Run the Prediction Fairness Explorer MicroSim Fullscreen

About This MicroSim

Bias in prediction is a systematic difference in how well a model's estimates fit different groups of learners, so that some are mis-estimated more often or in a consistent direction. Chapter 24 states the evaluation method: once data exists, compute the same prediction-error measure separately for each group and compare. A gap that the pooled average hides is a bias signal.

This chart applies that method to three groups: pointer users, keyboard users and learners on a shared device. Each group has a pair of bars:

  • Under-estimated (blue): the share of the group's learners the model predicted "not yet" who then passed a held-out check.
  • Over-estimated (orange): the share it predicted "mastered" who then failed the check.

Keeping the two directions apart matters. Chapter 24's first hypothesis is that a keyboard or screen-reader learner may never produce hover evidence, so the model could read less activity as less understanding. That would show up as under-estimation concentrated in one group. The dashed lines show the pooled rates over all learners.

All numbers are synthetic. They were invented to illustrate the method, and no real learner data has been collected through MicroSims. The three scenarios are:

Scenario What the synthetic data shows
Even error Every group sits close to the pooled rates
Hidden gap Keyboard users are under-estimated 37.5 percent of the time, yet the pooled rate is only 10.9 percent because the group is small
Small group Only 12 learners use a shared device; raising the threshold above 12 suppresses the group and its gap disappears from view

The Threshold mimics the Full LRS design's suppression rule, which hides any report cell built from fewer than the district's threshold, 10 students by default. Suppression protects privacy, but it also hides exactly the small groups a fairness check most needs to see, and that limit should be disclosed. The pooled lines still count the suppressed learners, so a reader who knows every count could subtract to recover the hidden group, which is why Chapter 25 also calls for complementary suppression.

A text table of the data, updated with every change, is available to screen readers, and the chart's accessible name summarizes the current values.

Learning objective: The learner will critique a mastery prediction by comparing its error across learner groups, and decide whether a pooled average hides a gap.

Bloom's taxonomy level: Evaluate (verb: critique)

How to Use

  1. Start with Even error. Compare each group's bars with the dashed pooled lines.
  2. Switch to Hidden gap. Read the pooled rates first, then the Keyboard users bars. Decide whether the model's pooled error is an acceptable summary.
  3. Turn Show pooled average off and on. Notice how easy it is to report only the pooled rate.
  4. Switch to Small group and drag Threshold from 10 to 13 or higher. The Shared device group is suppressed; read the message about what can no longer be seen.
  5. Hover over any bar to see the count behind the percentage. Press Reset to return to the starting view.

Iframe Embed Code

You can add this MicroSim to any web page by adding this to your HTML:

1
2
3
4
<iframe src="https://dmccreary.github.io/microsims/sims/prediction-fairness-explorer/main.html"
        height="602px"
        width="100%"
        scrolling="no"></iframe>

Lesson Plan

Audience

Teachers, instructional designers, learning-technology developers and learning-analytics practitioners (college undergraduate and professional development).

Duration

15-20 minutes

Prerequisites

  • Mastery prediction and held-out assessment (Chapters 18 and 19)
  • Bias in prediction and the suppression threshold (Chapters 21 and 24)

Activities

  1. Pooled first (3 min): In Hidden gap, cover the bars and read only the pooled rates. Write one sentence a dashboard might print about this model.
  2. Critique (5 min): Uncover the bars. Rewrite the sentence so it is honest about the Keyboard users group, and name the Chapter 24 hypothesis that could explain the direction of the error.
  3. Suppression trade-off (5 min): In Small group, raise the threshold until the group is hidden. Discuss who is protected and who is harmed when a gap disappears from a report.
  4. Disclosure (optional, 5 min): Draft the limits sentence a fidelity report should include when a group is too small to report.

Assessment

  • The learner states whether the pooled average hides a gap in each scenario, with the group, the direction and the size of the gap.
  • The learner explains why under- and over-estimation should be reported separately.
  • The learner explains what suppression hides and why its limit must be disclosed.

References

  1. Fairness (machine learning) - Wikipedia. Group fairness criteria that compare error rates across groups.
  2. Simpson's paradox - Wikipedia. How aggregating groups can hide or reverse what happens inside them.
  3. Statistical disclosure control - Wikipedia. Cell suppression and why complementary suppression is needed.
  4. Bar chart documentation - Chart.js. The grouped bar chart used here.