Accuracy ↑
Share of predictions whose class is correct. The point estimate is the posterior expectation of accuracy. Under class imbalance, read it together with AUPRC.
01
EVALUATION OVERVIEW
Overview
SEMI-SUPERVISED MODEL EVALUATION
Model evaluation.
A closer look.
Start from a few known labels and inspect candidate performance, relationships, and posterior structure.
OFFLINE · INTERACTIVE · TRACEABLEPoints are performance estimates. Thin lines are sampling intervals. A collapsed interval is drawn as a single point.
EVALUATION SETUP
Every conclusion starts from a setup.
Evaluation size, candidate models, and fit configuration.
Data composition
Click a model name to pin the highlight across charts. Click again to clear it. Tab and Enter work as well.
02
PERFORMANCE & RANKING
Performance and ranking
PERFORMANCE
Performance, metric by metric.
Compare candidates on overall classification, discrimination, positive-class retrieval, and probability calibration. Lower ECE is better. Higher is better for the other metrics.
AUC ↑
Area under the ROC curve: how well the model ranks positives ahead of negatives. Binary tasks use the positive-class probability. Multiclass tasks use macro one-vs-rest.
AUPRC ↑
Implemented as average precision, summarizing the precision–recall curve. Multiclass tasks use macro one-vs-rest. The value depends on the class balance.
ECE ↓
Expected calibration error over 10 equal-width bins, weighted by bin count. Binary tasks calibrate the positive-class probability. Multiclass tasks calibrate top-label confidence.
Performance matrix
Colored by relative performance within each columnAll four plots follow the input order M01, M02, and so on, matching the table rows. Purple marks the leading point estimate for that metric. Each horizontal axis is scaled on its own, so distances cannot be compared across plots.
Points are performance estimates. Thin lines are the central sampling interval. Accuracy uses the exact posterior expectation. AUC, AUPRC, and ECE use the mean of valid draws. Intervals come from latent-label draws under one fitted posterior. A narrower interval only means less variation from that source.
The table below keeps exact values, the number of valid draws, and the Monte Carlo standard error. That standard error measures numerical error in the sampling calculation. It is not the error of the performance estimate relative to the true value.
Per-model metrics · point estimates, sampling intervals, and Monte Carlo numerical error
SAMPLING & RANKING
One set of draws, for both values and ranks.
The same latent-label draws link metric distributions, model ranks, and pairwise gaps. Together they describe sampling variation.
If a sample has posterior P(class 1) = 0.7, each draw assigns it to class 1 with probability 70%. Across draws, the rank distribution counts how often each model occupies each rank. Pairwise sampling frequency counts how often the row model beats the column model. The repeated step is label sampling. Models are not retrained, and SSME is not refit.
Rank distribution
Pairwise sampling frequency
Rank-1 stability is a frequency among sampled evaluations, not the probability that a model is truly best. Top-2 uses the same tie rule.
How much do the metric values move across the same draws?
Metric sampling viewEach distribution collects one model's metric across draws. The paired comparison below uses the same draw for both A and B. An Accuracy delta of 0.002 means A is 0.2 percentage points higher than B on that draw. If every draw gives almost the same value, the distribution collapses to a point.
The outline is the sampling density, the thick line is the interquartile range, and the point is the median. When there is no visible variation, only a single point is shown.
PAIRWISE FOCUS
Compare two candidates
vs.
Delta is metric A minus metric B. For ECE, a negative value means A is better. For the other metrics, a positive value means A is better. The initial pair is the top two point estimates.
03
MODEL RELATIONSHIPS & CONTRIBUTION
Relationships and contribution
MODEL LANDSCAPE
Similar predictions, different views.
Model relationships are read directly from the input probabilities, without latent-label sampling.
Prediction correlation
Prediction agreement
MODEL CONTRIBUTION
What changes when one model is removed?
Drop each model's predictions in turn, refit SSME, and measure the effect on the joint estimate.
04
PREDICTION SPACE & POSTERIOR DIAGNOSTICS
Prediction space and diagnostics
PREDICTION SPACE
Inferred class, in two and three dimensions.
Each point is a sample. Position is the ALR-to-PCA projection of all model predictions. Color is the inferred class, not the true label. The 3D view adds the third principal component and does not refit SSME.
3D prediction space
Drag to rotate · scroll to zoom · arrow keys to rotate · + / − to zoomThe first three principal components of the same PCA, on a shared scale. Rotation only changes the viewing angle. Overlap in the projection does not mean the original predictions are identical.
POSTERIOR DIAGNOSTICS
How concentrated are the latent labels?
These uncertainty summaries include unlabeled samples only. Known labels stay fixed as one-hot and are left out.
Posterior confidence
Posterior entropy
The KDE uses Scott bandwidth and boundary reflection to display the empirical distribution. The smoothing width is not an estimation error. Near-constant data collapses to a point instead of a fabricated curve.