.png){kind=icon}
.png){kind=icon}
Return
Compositional feature augmentation for improving multi-class classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108519.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent studies on multi-class classification have made significant progress. However, many approaches still suffer from unsatisfactory accuracy, high computational cost, and insufficient class-specific feature representation. To address these issues, we propose a simple yet effective framework for multi-class classification, termed Compositional Feature Augmentation (CFA). By transforming each original feature into a class-wise posterior composition, CFA captures discriminative information in a model-independent and marginal learning manner. To enhance robustness, we introduce a voting mechanism that aggregates predictions from multiple augmented feature sets generated via random subsampling. This ensemble approach stabilizes results and mitigates noise in marginal learning. CFA is compatible with standard classifiers, including logistic regression, SVM, neural networks, and others. Extensive experiments on both structured and deep-embedded datasets show that CFA improves accuracy, especially when original features are unrefined, while maintaining competitive performance when strong embeddings are already available.
Journal
IF:
6.3
Papers:
7.8K
Citations:
3.0W