TBU Publications
Repository of TBU Publications

Depth estimation in half-space trees and feature range formulation for novelty detection

DSpace Repository

Show simple item record


dc.title Depth estimation in half-space trees and feature range formulation for novelty detection en
dc.contributor.author Ulrich, Adam
dc.contributor.author Krňávek, Jan
dc.contributor.author Hrabec, Dušan
dc.contributor.author Šenkeřík, Roman
dc.relation.ispartof Information Sciences
dc.identifier.issn 0020-0255 Scopus Sources, Sherpa/RoMEO, JCR
dc.identifier.issn 1872-6291 Scopus Sources, Sherpa/RoMEO, JCR
dc.date.issued 2026
utb.relation.volume 756
dc.type article
dc.language.iso en
dc.publisher Elsevier Inc.
dc.identifier.doi 10.1016/j.ins.2026.123823
dc.relation.uri https://www.sciencedirect.com/science/article/pii/S0020025526007541
dc.subject Novelty detection en
dc.subject Anomaly detection en
dc.subject Half-space trees en
dc.subject Depth estimation en
dc.subject Density-aware models en
dc.subject Feature range estimation en
dc.description.abstract Tree-based methods such as Isolation Forest (iForest) and Half-Space Trees (HS-Trees) are widely used for unsupervised anomaly and novelty detection due to their efficiency, scalability, and ability to operate without labelled anomalies. However, the classical depth estimator c(n), derived for iForest under random partitioning assumptions, is not theoretically consistent with deterministic half-space splits used in HS-Trees. We analyse an idealised generative model of HS-Trees for novelty detection and derive a theoretically grounded formulation of the expected isolation depth, leading to a modified estimator c′(n). We provide both an exact recursive expression and a computationally efficient closed-form approximation of c′(n), and validate it through analytical arguments, Monte Carlo simulations and real dataset (ADBench) evaluations. In the second part, we investigate the problem of initial feature-range specification, a critical yet often overlooked component of novelty detection, and systematically evaluate several statistically motivated range initialisation strategies. Extensive experiments on the ADBench benchmark, including ablation studies and nonparametric statistical tests, show that the proposed depth formulation preserves detection performance relative to established baselines. Overall, the paper provides a theoretical basis for depth scaling in HS-Trees and clarifies how range design affects tree-based novelty detection. en
utb.faculty Faculty of Applied Informatics
dc.identifier.uri http://hdl.handle.net/10563/1012881
utb.identifier.obdid 43887423
utb.identifier.scopus 2-s2.0-105043628327
utb.identifier.wok 001813785800001
utb.identifier.coden ISIJB
utb.source j-scopus
dc.date.accessioned 2026-11-20T08:16:39Z
dc.date.available 2026-11-20T08:16:39Z
dc.description.sponsorship This work was supported by the Internal Research Project of the Faculty of Applied Informatics, Tomas Bata University in Zl\u00EDn, No. RVO/FAI/2024/004 and No. IGA/CebiaTech/2023/004.
dc.rights Attribution 4.0 International
dc.rights.uri http://creativecommons.org/licenses/by/4.0/
dc.rights.access openAccess
utb.contributor.internalauthor Ulrich, Adam
utb.contributor.internalauthor Krňávek, Jan
utb.contributor.internalauthor Hrabec, Dušan
utb.contributor.internalauthor Šenkeřík, Roman
utb.fulltext.sponsorship This work was supported by the Internal Research Project of the Faculty of Applied Informatics, Tomas Bata University in Zlín, No. RVO/FAI/2024/004 and No. IGA/CebiaTech/2023/004.
utb.fulltext.projects RVO/FAI/2024/004
utb.fulltext.projects IGA/CebiaTech/2023/004

Files in this item

Files Size Format View

There are no files associated with this item.

Show simple item record

Attribution 4.0 International Except where otherwise noted, this item's license is described as Attribution 4.0 International