GEMEX Playground

Geodesic Sensitivity Field (GSF) measures how strongly each feature pushes the model's prediction toward or against the predicted class — but unlike SHAP or LIME, it does this by integrating directional sensitivity along the actual curved geodesic path on the statistical manifold, not along a flat straight line.

Green bar = feature supports the predicted class. Red bar = feature works against it. Longer bar = stronger influence. The shaded band on each bar is a curvature-weighted geometric uncertainty — it tells you how reliably the model uses that feature in this region of its decision surface. Wide band = geometrically complex region, interpret with care. This uncertainty measure does not exist in SHAP or LIME.

Features are sorted by absolute importance. Rescaling a feature (e.g. kg → lbs) never changes its GSF score — this reparametrisation-invariance is a mathematical guarantee unique to GEMEX.

Green = supports prediction  ·  Red = opposes  ·  Band = curvature uncertainty

Geodesic Journey Chart — two aligned panels showing what happens step-by-step as GEMEX walks the curved statistical manifold from baseline to instance.

Top — Local Manifold Stretch: how much distance is covered on the manifold at each RK4 integration step. Tall bars = the model's decision surface curves sharply at that step (a nearby decision boundary). The ⚡ gold peak marks the most geometrically complex step, with a gold vertical line connecting both panels. Three shaded zones — Baseline / Traversal / Instance — divide the path into thirds.

Bottom — Prediction Probability: how the model's confidence in the predicted class evolves along the path. The blue dot is the baseline; the green dot is the instance. The red dashed line at p = 0.5 is the decision boundary — if the probability curve crosses it, the model switched class during the explanation. The gold dot marks probability at the peak-curvature step.

Parallel Transport Interaction (PTI) quantifies nonlinear pairwise dependencies between features using the concept of holonomy from differential geometry: it measures the rotation accumulated when a feature's attribution vector is transported around a closed loop on the statistical manifold.

A non-zero PTI(i, j) means features i and j interact in the model's geometry in a way that cannot be decomposed into individual effects — the joint influence is genuinely irreducible. This captures interactions that SHAP interaction values, LIME coefficients, and all additive methods fundamentally miss, because those methods assume the attribution surface can be linearly decomposed.

Gold bar = synergistic: the two features amplify each other's contribution when acting together. Purple bar = antagonistic: they partially suppress each other. Longer bar = stronger geometric entanglement. Only interactions with |PTI| > 0.02 are shown. Requires interaction order ≥ 2.

Feature Attention Sequence (FAS) tracks which feature the geodesic path pays attention to at each integration step — recording the model's geometric reasoning process in the order it unfolds, from baseline to final prediction.

At each step, GEMEX finds which feature direction the statistical manifold is most stretched along (the leading eigenvector of the Fisher Information Matrix). The fraction of steps where each feature dominates is its dwell time, shown as a percentage bar.

Blue bars = features dominant in the early path (near baseline — the model's initial reasoning). Gold bars = features dominant in the late path (near the decision boundary — final refinement). Unlabelled bars = features that appear in both phases or neither.

This temporal dimension of explanation — in what order does the model reason about features? — does not exist in any other XAI method. The number of attention switches shown at the top tells you how many times the model's geometric focus changes along the path. Requires order ≥ 2.

Bias Trap Detector (BTD) is a geometric audit that identifies features where the model pays disproportionate attention relative to the actual influence they have on the prediction — a warning signal for potential confounding, over-reliance, or spurious correlation. This diagnostic has no equivalent in SHAP, LIME, or GradCAM.

Each bar combines three independent geometric bias signals:
Red — HAT (Holonomy Asymmetry Test): the feature generates asymmetric curvature loops, meaning the model responds differently to it depending on context rather than its value alone. This is the classic signature of a confounder.
Gold — MCA (Manifold Curvature Asymmetry): the Fisher Information Matrix is unusually sensitive in this feature's direction without proportionate attribution — the model is geometrically over-reliant on it.
Purple — GDI (Geodesic Dominance Inconsistency): high attention dwell time combined with low final GSF score — the model "thinks about" this feature a lot but doesn't translate that into output.

Features labelled HIGH should be reviewed before deployment, particularly if they correspond to age, sex, ethnicity, or other sensitive attributes. Requires order ≥ 2.

Riemannian Curvature Triplet (RCT) measures three-way feature interactions using the Riemann curvature tensor projected onto feature triplets — the only XAI method that mathematically isolates how a third feature modulates the interaction between two others.

Each row shows a triplet (Feature A × Feature B) → Feature C, meaning: the way A and B interact changes depending on the value of C. This three-way modulation is irreducible — it cannot be reconstructed from any combination of pairwise PTI values or individual GSF scores.

Teal bar (positive RCT) = synergistic modulation: A and B amplify each other's effect on C's contribution. Orange bar (negative RCT) = antagonistic modulation: A and B suppress each other's effect on C. Longer bar = stronger three-way entanglement.

Requires interaction order = 3. Computational cost scales as O(n³) — for 13 features this means 286 unique tensor entries. No other XAI library computes this quantity.