Multi-Model Fusion for Anomaly Detection
Artykuł w czasopiśmie
MNiSW
100
Lista 2024
| Status: | |
| Autorzy: | Gałka Łukasz, Kozieł Grzegorz, Dziuba-Kozieł Marta |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 14 |
| Strony: | 143714 - 143734 |
| Impact Factor: | 4,2 |
| Scopus® Cytowania: | 0 |
| Bazy: | Scopus |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | TAK |
| Licencja: | |
| Sposób udostępnienia: | Witryna wydawcy |
| Wersja tekstu: | Ostateczna wersja opublikowana |
| Czas opublikowania: | W momencie opublikowania |
| Data opublikowania w OA: | 14 września 2026 |
| Abstrakty: | angielski |
| Anomaly detection is a crucial challenge for modern information systems. It helps in data cleansing and identifying outliers with unique traits. However, unsupervised detectors often vary greatly in performance across datasets. They are also sensitive to the choice of algorithm and its hyperparameters. As a result, relying on a single model can be risky in practice. To address this issue, we investigate a multi-model fusion strategy. The goal is to improve robustness under heterogeneous detector behavior. Our approach combines heterogeneous detectors in a plug-and-play manner without training an additional meta-model. Several fusion rules are adapted from prior ensemble and information-fusion studies because they are model-agnostic and require no additional training. Therefore, we propose fusion techniques including majority voting, probability averaging, the ordered weighted averaging operator, the Choquet integral, and the Takagi-Sugeno model. Experiments use models from the widely adopted PyOD library, testing combinations of two, three, and four models across 28 real-world datasets. Detector training and fusion-score generation are label-free. However, ground-truth labels are used for operating threshold and best configuration selection during evaluation. Therefore, the reported accuracy and F1 score results should be interpreted as oracle upper-bound estimates. Under this protocol, average accuracy improves by about 8%, and the F1 score by over 13%. Statistical tests at α=0.05 further support the potential effectiveness of model fusion over single-model methods. The framework operates fully in parallel, and the additional computational cost from the fusion stage is minimal compared to training and evaluating the base models. |
