Theory-driven anomalous sound detection at Aalborg University: AUC bounds, pseudo-anomalies, self-optimizing ensembles
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Insufficient data: the snapshot provides no aggregate publication counts, venue record, or output volume, so this axis — which would measure publishing productivity and volume over time — is left unscored rather than estimated.
Impact—
No citation counts, h-index, or uptake metrics are present in the input data, so this axis — which would measure how widely the work is cited and adopted by others — cannot be scored and is reported as missing.
Focus0/100
The radar focus value is recorded as 0, and the raw field distribution confirms an extremely narrow footprint: the two Evolve-reviewed papers split evenly (0.5/0.5) between two near-identical labels — machine learning for audio anomaly detection and machine learning for anomaly detection / anomalous sound detection. Within this reviewed sample, the work concentrates almost entirely on unsupervised anomaly detection theory and methods for sound data.
Collaboration28/100
The collaboration score is 28, computed from a small circle of 3 distinct coauthors in the reviewed data. Kevin Wilkinghoff is the recurring partner, listed twice among frequent coauthors, and on both 2026 catalog papers Zheng-Hua Tan takes the last-author position without corresponding authorship — a pattern consistent with senior supervisory involvement in a compact core team rather than a large consortium.
Activity—
The snapshot covers only two 2026 arXiv entries, with no multi-year publication cadence available, so this axis — which would measure recent publishing tempo and ongoing research engagement — lacks the underlying data needed for scoring.
Reproducibility55/100
Reproducibility scores 55 (reproMean 55, computed over the 2 papers with readiness assessments), both rated medium. The work is anchored on the public DCASE anomalous sound detection benchmark, with explicit, checkable theoretical bounds such as AUC lower bounds for kNN detectors; however, the best-performing pseudo-anomaly constructions depend on metadata availability, which tempers out-of-the-box replication.
Research summary
Zheng-Hua Tan is a researcher at Aalborg University whose Evolve-reviewed work centers on machine learning for anomaly detection, and specifically on anomalous sound detection. The current catalog holds two 2026 arXiv papers, on both of which Tan is last author (non-corresponding), working with recurring co-author Kevin Wilkinghoff inside a small circle of three distinct coauthors. The thread is unusually theory-forward for the field: one paper derives an AUC lower bound for kNN detectors and asks whether detection performance can be predicted from embedding-space geometry, using pseudo-anomaly probes for anomaly-free model selection on DCASE — though the strongest constructions need metadata. A companion paper recasts AUC bounds as a differentiable objective for anomaly-free self-optimization, tuning ensembles without anomaly labels; it reports strong DCASE gains under LDN scoring and no gain under plain NN scoring. Both papers carry medium reproducibility readiness (reproMean 55), reflecting reliance on the public DCASE benchmark alongside metadata-dependent components. The radar records an extremely narrow topical footprint — a focus value of 0 with the field distribution split evenly between two near-identical anomaly-detection labels — and a collaboration score of 28, while output, impact, and activity axes remain unscored for lack of data. This profile is name-resolved from public bibliographic data; ORCID and country are unconfirmed in the snapshot.
Collaboration pattern
The reviewed sample shows a deliberately small collaboration footprint: 3 distinct coauthors, with Kevin Wilkinghoff the dominant recurring partner (listed twice among frequent coauthors). On both 2026 arXiv papers Zheng-Hua Tan takes the last-author position without corresponding authorship, a signature typical of a senior researcher guiding a focused small team rather than leading large consortia. No collaboration data beyond this circle is present in the snapshot.
Research keywords
Highlights
AUC lower bounds for kNN anomaly detectors
Pseudo-anomaly probes for label-free model selection on DCASE
Self-optimizing ensembles without anomaly labels
Timeline
- 2026An arXiv preprint asks whether anomaly detection performance can be predicted from embedding-space geometry, deriving an AUC lower bound for kNN detectors; the Evolve review notes pseudo-anomaly probes enable anomaly-free model selection on DCASE, with the best constructions needing metadata.
- 2026A follow-up arXiv preprint recasts AUC bounds as a differentiable objective for anomaly-free self-optimization, tuning ASD ensembles with strong DCASE gains under LDN scoring but no gain under plain NN scoring.
- 2026Both 2026 entries enter the Evolve review catalog with Zheng-Hua Tan as last author (non-corresponding), each rated medium for reproducibility readiness (reproMean 55 across 2 papers).
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric figures and radar scores cover only the two Evolve-reviewed papers in this catalog; no full publication record is included.
- Output, impact, and activity axes are unscored: no citation counts, h-index, venue, or multi-year productivity data were provided.
- ORCID and country are not resolved; the profile is name-resolved from public bibliographic data only.
- Both catalog entries are 2026 arXiv preprints; peer-reviewed publication status is not captured.