.png){kind=icon}
.png){kind=icon}
Return
Stacked fuzzy envelope consistency imbalanced ensemble classification method
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.126033.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Ensemble methods are commonly employed to address Class imbalance(CI) issue due to their effectiveness. However, existing imbalanced ensemble (IE) approaches are typically applied directly to the original subsets during classifier training. Since the original subsets suffer from low separability and diversity in many cases, the accuracy is limited. To address the issue, a new IE approach called the stacked fuzzy envelope consistency imbalanced ensemble method (SFECEM) is proposed. First, balanced subsets are obtained by the subsets generation model (BSG). Then, a multilayer fuzzy envelope instance generation mechanism with local-global structure consistency by using the stacking model (SFCLGSC) is proposed to obtained envelope subsets. In the SFCLGSC, the multilayer fuzzy clustering is performed simultaneously with the structure consistency maintenance in the lower dimensional space and the global and local structures are included. Finally, the original and envelope instances in the lower dimensional space are trained and fused by a sparse weight fusion mechanism (SWFM). Some experiments were conducted using multiple public datasets and over ten well-established relevant algorithms were selected for verification. The results indicate that the proposed algorithm significantly outperforms other IE methods. Specifically, compared to the state-of-the-art (SOTA) methods, SFECEM shows improvements of 9.54% in AUC, 14.07% in F-M, and 9.55% in G-M. The key contributions of this article are: (a) designing the SFCLGSC model to generate envelope instances; (b) optimizing clustering, global and local distributions, and instance dimension jointly; and (c) proposing an IE algorithm that more effectively addresses the CI issue.
Keywords:
Ensemble method
Class imbalance
Fuzzy clustering
Local-global structure consistency
Envelope transformation
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
View more organizations