Logo do repositório
 
Publicação

Label-Free Visual Concept Drift Detection via Classifier Two-Sample Tests and Characteristic Function Embeddings

2026Artigo científico Acesso aberto
Ver registo simples
dc.contributor.authorHovakimyan, Gurgen
dc.contributor.authorBravo, Jorge Miguel
dc.contributor.institutionInformation Management Research Center (MagIC) - NOVA Information Management School
dc.contributor.institutionNOVA Information Management School (NOVA IMS)
dc.date.accessioned2026-08-21T09:11:01Z
dc.date.available2026-08-21T09:11:01Z
dc.date.issued2026
dc.descriptionHovakimyan, G., & Bravo, J. M. (2026). Label-Free Visual Concept Drift Detection via Classifier Two-Sample Tests and Characteristic Function Embeddings. Journal of Soft Computing and Decision Analytics, 4, 135-55. https://doi.org/10.31181/jscda41202686
dc.description.abstractDeploying deep neural networks in non-stationary environments exposes them to concept drift, which silently degrades predictive performance over time. Traditional statistical two-sample tests and distance-based heuristics frequently fail to detect natural, semantic shifts in high-dimensional visual streams. To address this critical blind spot, we propose a real-time, label-free concept drift detection framework that couples the Classifier Two-Sample Test (C2ST) with a novel Characteristic Function Embedding (CFE) bottleneck. Operating directly on deep convolutional representations, this architecture compresses features to resolve dimensionality constraints and ensure highly efficient downstream evaluation. We systematically benchmark our sliding-window framework against classical and modern baselines under complex, real-world domain and subpopulation shifts, as well as controlled synthetic perturbations. Our empirical findings demonstrate that distance-based baselines severely degrade under semantic shift, whereas the proposed C2ST-CFE framework retains robust discriminative power. Furthermore, the framework successfully isolates ultra-low intensity synthetic drift while maintaining a calibrated false-positive rate on stable, in-distribution streams. These results confirm that the proposed architecture provides a mathematically grounded, highly responsive, and computationally efficient mechanism for safe machine learning deployment in real-world visual applications.en
dc.description.versionpublishersversion
dc.description.versionpublished
dc.format.extent21
dc.format.extent2099269
dc.identifier.doi10.31181/jscda41202686
dc.identifier.issn3009-3481
dc.identifier.otherPURE: 168093055
dc.identifier.otherPURE UUID: a94c6f11-b9c1-46d0-971b-8ca7c612a62b
dc.identifier.otherScopus: 105047869172
dc.identifier.otherORCID: /0000-0002-7389-5103/work/224537192
dc.identifier.urihttp://hdl.handle.net/10362/205612
dc.identifier.urlhttps://www.scopus.com/pages/publications/105047869172
dc.language.isoeng
dc.peerreviewedyes
dc.relationhttps://doi.org/10.54499/UID/04152/2025
dc.relationhttps://doi.org/10.54499/UID/PRR/04152/2025
dc.relationhttps://doi.org/10.54499/UID/00315/2025
dc.subjectConcept drift
dc.subjectClassifier two-sample test
dc.subjectSemantic shift
dc.subjectDeep representations
dc.subjectLabel-free monitoring
dc.subjectArtificial Intelligence
dc.subjectDecision Sciences (miscellaneous)
dc.subjectManagement Science and Operations Research
dc.subjectApplied Mathematics
dc.titleLabel-Free Visual Concept Drift Detection via Classifier Two-Sample Tests and Characteristic Function Embeddingsen
dc.typejournal article
degois.publication.firstPage135
degois.publication.lastPage55
degois.publication.titleJournal of Soft Computing and Decision Analytics
degois.publication.volume4
dspace.entity.typePublication
rcaap.rightsopenAccess

Ficheiros

Principais
A mostrar 1 - 1 de 1
Miniatura
Nome:
Label-Free_Visual_Concept_Drift_Detection.pdf
Tamanho:
2 MB
Formato:
Adobe Portable Document Format
Ver/Abrir

Coleções

NIMS: MagIC - Artigos em revista internacional com arbitragem científica (Peer-Review articles in international journals)