.png){kind=icon}
Return
Attribute Grouping-Based Outlier Detection Using Dynamic Dimensional Embedding Representations
.png){kind=icon}
PRE
AI
DOI：10.1002/cpe.70463.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Attribute grouping serves as one of the effective steps in high-dimensional outlier detection. However, existing outlier detection methods merely capture the local coupling between pairs of attribute values within each attribute group, and overlook the global coupling among combinations of multiple attribute values, thereby limiting the effectiveness of outlier detection. In this paper, a novel attribute grouping-based outlier detection method for categorical data is proposed by using dynamic dimensional embedding representations to characterize the global coupling among all variables. First, within each attribute group, we dynamically construct the embedding dimensions of all attributes to avoid redundant representations caused by static dimensions. Second, Multi-Head Self-Attention is used to construct the dynamic dimensional embedding of any data object, effectively characterizing the global coupling. Then, the reconstruction error derived from the dynamic dimensional embedding is employed to quantify the outlier degree of data objects. Based on this, we propose an attribute grouping-based outlier detection method, in which dynamic dimensional embedding representations are used for each attribute group. In the end, experimental results on the UCI and synthetic datasets validate that the algorithm has good outlier detection performance. Importantly, compared with the competitive methods, the algorithm bolsters the AUC index and the detection efficiency by an average of 7.76% and 48.49%, respectively.
Keywords:
attribute grouping
dynamic dimensional embedding
Multi-Head Self-Attention
outlier detection
reconstruction error
Journal
C
IF:
1.5
Papers:
473
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSave
View more