Hierarchical One-Class Data Description via Probabilistic Granular-ball Computing

Lifeng Shen, Ruiwen Liu, Liang Peng, Shuyin Xia, Yi Wang, YI LIU

Keywords: One-class Classification; Data Description; Anomaly Detection; Probabilistic Granular-ball

Abstract:

One-class data description aims to model the distribution of target data by constructing a compact representation of the target class. This approach is widely applied in tasks like anomaly detection, where the objective is to differentiate target data from outliers. Traditional methods typically rely on single-sphere or pre-defined multi-sphere representations. However, these simplistic assumptions often fail to capture the anisotropic structures and intricate patterns present in real-world data, limiting their effectiveness in representing distributions across multiple scales.
To address these limitations, we propose Probabilistic Granular-ball Computing (PGBC), a hierarchical framework for one-class data description. PGBC uses ellipsoidal granular-balls to align with the anisotropic geometry of data and recursively refines them through statistical splitting, achieving precise and adaptive data representation. Moreover, PGBC approximates a hierarchical Gaussian mixture model by aggregating data description scores via granular-ball distribution entropy at each layer. This enables PGBC to capture data patterns at multiple levels of granularity, modeling both global structures and fine local variations.
Extensive experiments on benchmark datasets demonstrate that PGBC consistently outperforms related strong baselines, offering superior accuracy for hierarchical one-class data description while maintaining a low false positive rate.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 7514