.png){kind=icon}
.png){kind=icon}
Return
Density-Based Online Concept Evolution Detection in Streaming Features
.png){kind=icon}
PRE
AI
DOI：10.1109/TETCI.2025.3597301.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In practical applications, data are often dynamic, generated, and collected continuously. There are two types of streaming data: streaming samples and streaming features. Online learning from streaming data faces two primary challenges: infinite data length and changing concepts. Much research focuses on concept changes in sample streams, such as concept drift. In contrast, existing methods for streaming feature learning predominantly emphasize feature selection or online classification. However, concept evolution also occurs in feature streams. This paper introduces a novel method for detecting concept evolution in feature streams, termed D-CED-FS. This method comprises an adaptive sliding window, an improved parameter-free clustering algorithm, a density-based concept evolution prediction mechanism, a similarity-based concept evaluation method, and a visualization tool using weighted bipartite graphs. The adaptive sliding window module adjusts the size of detection windows in real time based on concept evolution in adjacent windows. The efficient adaptive clustering module, enhanced through natural neighbor methods, eliminates the need for parameter setting and can automatically generate the optimal concept set for each feature window to accommodate feature streams. The concept evolution prediction module reduces repeated clustering for each window, significantly enhancing the algorithm's detection efficiency. The visualization tool, employing a weighted bipartite graph, intuitively illustrates various types of detected concept evolution. Extensive experiments on synthetic and high-dimensional datasets demonstrate the effectiveness of the clustering algorithm and D-CED-FS's capability to detect concept evolution in feature streams.
Keywords:
Streaming features
concept evolution detecting
density peaks clustering
Journal
I
IF:
6.5
Papers:
1.4K
Citations:
4.5K