ML Ethics — 3D Fairness Landscape

Genuine WebGL companion: a real algorithmic-fairness simulation — two demographic groups scored by a live classifier in 3D feature space, with real disparate-impact, accuracy and audit statistics (the 2D original only mapped its sliders straight onto decorative gauges, so this builds the actual fairness mechanic the title describes).

Group A positive rate

—%

Group B positive rate

—%

Disparate impact

—

Accuracy vs ground truth

—%

Audited points

Audit-sample D.I.

—

Interpretation (four-fifths rule)

—

Historical data bias

Bias level (group B disadvantaged)0.3

Fairness intervention

Equalized-threshold adjustment0.0

Transparency

Explain decision surface & weights0.6

Accountability

Audit coverage0.2

Playback

Group A (baseline)

Group B (data-biased)

Classified positive (ring)

Currently audited

ML Ethics — 3D Algorithmic Fairness Landscape

The 2D original ("Advanced Machine Learning Ethics Simulation") only mapped its four sliders — fairness, transparency, accountability, bias — linearly onto four decorative percentage gauges and a radar chart; no classifier, dataset or fairness metric was actually computed. This 3D companion builds the real mechanic the title implies: a synthetic population of two demographic groups is scored by a genuine logistic classifier in 3D feature space, and the fairness statistics displayed are computed from the actual classification outcomes, not read off a slider.

Each point is one individual, positioned by their observed qualification score (x) and risk score (z), with height (y) equal to the classifier's real decision score sigmoid(0.8·qualification_norm − 0.6·risk_norm). The Bias level slider controls how much Group B's observed data is historically distorted relative to its true (equal) underlying ability distribution — a literal model of proxy discrimination in training data. The Fairness intervention slider applies a real equalized-threshold correction, lowering Group B's decision threshold to counteract that recorded bias. Transparency reveals the classifier's actual weight vectors and per-group decision-threshold planes; Accountability draws a genuine random audit subsample and recomputes its disparate-impact ratio independently of the full population.

The readouts — Group A/B positive rates, the disparate-impact ratio (the real "four-fifths rule" used in US employment-discrimination law), accuracy against an independent ground-truth label, and the audit-sample disparate impact — are all recalculated live from the 3D scene's actual point data whenever a slider moves or the dataset is regenerated. Drag to orbit, scroll or pinch to zoom.