.png){kind=icon}
.png){kind=icon}
Return
Semi-supervised feature selection with concept factorization and robust label learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.112317.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Feature selection (FS) is essential for improving model performance in high-dimensional data by identifying the most relevant features. Concept Factorization (CF), building on Non-negative Matrix Factorization (NMF), is valued for revealing meaningful data structure and producing interpretable concept vectors. However, existing CF-based FS methods are typically unsupervised and do not leverage label information, leading to a bias toward high-variance features. This bias can result in the omission of low-variance features that may be highly discriminative, ultimately reducing the effectiveness of FS and compromising model performance, especially in tasks where subtle or rare patterns are important. To address these limitations, this paper proposes SCFLR, a novel semi-supervised FS method that combines CF with robust label learning. SCFLR establishes the CF framework based on the feature space by expressing each concept vector as a conic combination of the feature vectors, thereby leveraging both the underlying data structure and available label information to select a more informative and balanced set of features. To this end, SCFLR defines a linear regression-based loss function derived from the generated concept vectors to leverage information from labeled data. This loss function is further enhanced through a label learning framework based on the L2,1-norm to ensure a robust label approximation. SCFLR also utilizes the dual-graph regularization to maintain the local geometric structures in both feature and data spaces. In order to tackle the optimization problem of SCFLR, an efficient algorithm, with proof of its convergence, is introduced. Finally, the experimental validation of the SCFLR method on multiple datasets highlights its effectiveness and superior performance compared to other FS methods.
Keywords:
feature selection
concept factorization
semi-supervised learning
label learning
robust regression
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
View more organizations