Back to search

University of Patras

Unsupervised outlier detection: new approaches for robust ensembles and interpretability

Abstract

dc:description

Outlier detection (OD) is crucial for identifying anomalies in fields like fraud detection, network security, and medical diagnostics. However, model selection and interpretability remain significant challenges in unsupervised settings. This thesis addresses these challenges by introducing methodologies for improved model selection and interpretability. First, a meta-learning ensemble framework is introduced for improved model selection. This framework treats outlier scores from multiple detectors as features and employs unsupervised feature selection to identify the most informative detectors for a given dataset. The selected detectors are then combined through an ensemble voting approach. Second, a taxonomy of interpretability methods is presented. This taxonomy categorizes existing techniques based on their scope (local vs. global), model dependency (model-agnostic vs. model-specific), and whether they are applied during or after model training (intrinsic vs. post hoc). The taxonomy also considers applicability across diverse data types, including tabular, text, image, and graph data. Finally, an unsupervised OD method is proposed that leverages both the density estimation capabilities of normalizing flows and the interpretability of decision trees. The method uses normalizing flows to generate pseudo-labels, which then train a decision tree classifier. This provides clear, rule-based classification boundaries for outlier detection, informed by the underlying data density. These contributions collectively facilitate the development of unsupervised OD systems with improved model selection, enhanced interpretability, and a clearer understanding of the underlying decision-making process, ultimately supporting the deployment of these systems in critical real-world applications.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Papastefanopoulos, Vasileios
  • Παπαστεφανόπουλος, Βασίλειος

Subjects

dc:subject × 24

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/59692

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Papastefanopoulos, Vasileios; Παπαστεφανόπουλος, Βασίλειος. Unsupervised outlier detection: new approaches for robust ensembles and interpretability. University of Patras, 2025. http://hdl.handle.net/10442/hedi/59692