Unsupervised ranking of anomaly detection models via cluster-holdout pseudo-labels.

For each detector D, cluster the unlabeled data, hide a subset S of clusters, train D on the complement, and score points in S as pseudo-anomalies. Then the ranking of detectors by pseudo-AUC should correlate with their ranking by true AUC on real anomalies.

Datasets from ADBench (public benchmarks with real labels). Labels are used only to compute the ground-truth ranking; the pseudo-ranking pipeline never sees them. For each dataset we produce a rank correlation (Spearman, Kendall) between pseudo-rank and true-rank, plus top-1 / top-3 hit rates.

• src/adrank/pipeline.py — datasets, detectors, clustering, pseudo-eval, aggregation
• scripts/fetch_adbench.py — downloads ADBench Classical .npz files into data/adbench/
• scripts/run_v1.py — main experiment runner (parallel across dataset x subset)
• scripts/aggregate_report.py — Borda / mean-AUC / var-weighted aggregation + correlations
• scripts/smoke_test.py — 5-dataset synthetic smoke test (~30 s)

• Latent embedding: PCA to 16 dims (or d if d < 16).
• Clustering: MiniBatchKMeans with K ∈ {30, 50} (sweep).
• Cluster-composite anomaly-likeness score: 1/3 * (1 - size/N) + 1/3 * mean-dist-to-k-nearest-centroids + 1/3 * (1 - local-density), each min-max normalized within dataset.
• Subset draws: M = 20 per (dataset, K). Each subset targets ~5% of N as pseudo-anomalies, drawn from the top-quartile of the composite score. Held-out normal fold is 20% of the training complement.
• Detectors: IForest, LOF, KNN, OCSVM, ECOD, COPOD, HBOS, PCA, CBLOF, LODA with PyOD defaults, fixed seed.
• Aggregations reported: Borda-over-ranks, mean pseudo-AUC, variance-weighted mean.

Mean Spearman ρ ≥ 0.6 across ≥ 30 datasets, and top-1 hit-rate ≥ 40% (random baseline ≈ 1/10 = 10%).

1. Scrambled pseudo-scores → ρ near 0.
2. Trivial "all-same" pseudo-score baseline (mean intra-cluster density, detector-independent) should NOT match the pseudo-ranking's correlation. If it does, the signal is cluster geometry, not detector behavior.
3. Random cluster-subset selection (uniform, not composite-weighted) should give strictly lower correlation than the composite-weighted selection — otherwise selection does not matter and we defend the simpler variant.