.png){kind=icon}
.png){kind=icon}
Return
Fuzzy granule kernel density estimation for outlier detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.114168.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Outlier detection is a fundamental task in data mining for identifying rare instances that deviate from the norm. Its effectiveness is often hindered by datasets containing a heterogeneous mix of attributes, which introduce substantial uncertainty and blur the boundaries between concepts. Traditional detection methods that rely on crisp distance or density metrics struggle in such settings. Fuzzy rough set (FRS) theory, by leveraging information granulation to explicitly model this uncertainty, is particularly well suited to address these challenges. Nevertheless, existing FRS-based approaches exhibit a notable limitation as they primarily capture the relational structure among samples while often overlooking the informative role of local density distributions, thereby reducing their ability to detect subtle, locally sparse outliers. To bridge this gap, we propose fuzzy granule kernel density estimation for outlier detection (FGDOD), a novel algorithm that integrates fuzzy-granule representation with kernel density estimation in a unified framework. By employing a hybrid fuzzy similarity relation with an adaptive radius, FGDOD first constructs robust fuzzy granules for heterogeneous data. A kernel-based density quantification is then used to characterize the local distributional properties of these granules, and a density-weighted fusion strategy aggregates multi-granularity information to produce more discriminative outlier scores. Extensive experiments on 20 benchmark datasets demonstrate that FGDOD consistently outperforms mainstream detection algorithms, demonstrating superior effectiveness, robustness, and generalization capability across heterogeneous data types.
Keywords:
outlier detection
fuzzy rough set
kernel density estimation
heterogeneous data
outlier scoring
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W