MSTIF-IFS: Isolation forest based on minimal spanning tree and intuitionistic fuzzy sets
Artykuł w czasopiśmie
MNiSW
140
Lista 2024
| Status: | |
| Autorzy: | Gałka Łukasz |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 178 |
| Numer artykułu: | 113507 |
| Strony: | 1 - 12 |
| Impact Factor: | 7,6 |
| Efekt badań statutowych | NIE |
| Finansowanie: | This work was supported by the grant FD-20/IT-3/047. |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | NIE |
| Abstrakty: | angielski |
| Unsupervised anomaly detection represents one of the most thoroughly investigated and broadly utilized areas within machine learning. Among the most effective approaches in this domain is the Minimal Spanning Tree-Based Isolation Forest (MSTIF). This method constructs isolation trees by iteratively merging elements using Kruskal’s algorithm. This study introduces a novel enhancement of MSTIF by extracting structural parameters during the training phase and using them directly in the scoring process. The extracted parameters characterize the structure of isolation trees and enable a more precise, tree-structure-aware evaluation. Additionally, the proposed improvement integrates an innovative evaluation mechanism grounded in the theory of intuitionistic fuzzy sets (IFS). It estimates anomaly membership, non-membership, and associated uncertainty. The IFS-based measures are then used to derive the final anomaly score. Furthermore, an interpretability mechanism is developed to support score analysis in unsupervised settings by explicitly quantifying certainty and uncertainty. The proposed method demonstrates high detection performance, validated through experiments on 30 real-world anomaly detection datasets. Performance is assessed using evaluation metrics such as PR AUC and ROC AUC. Comparative analyses are performed against twelve recently introduced state-of-the-art algorithms. In all comparisons, the proposed approach consistently achieves statistically significant improvements in detection quality, underscoring its strong effectiveness. |
