.png){kind=icon}
.png){kind=icon}
Return
Imbalance multi-label data learning with label specific features
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2022.09.085.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
A class of machine learning problem where each instance may either belong to one or more than one class simultaneously is known as Multi-label classification problem. A well-known paradigm for multi-label classification is binary relevance (BR) where multi-label classification is decomposed into binary (one -vs-rest) classification subproblems, one for each label. However, it has a number of drawbacks. First, each binary classifier can be affected by the class-imbalance issue. Second, it does not take label correlations into account. Third, it has an inconsistency problem i.e. training instances with multi-label characteristics are considered both as positive as well as negative instance simultaneously. We attempt to resolve these problems in our proposed formulation. The problem of class imbalance is addressed by applying different weights to positive and negative examples depending on the class distribution, the inconsistency issue is addressed by selecting label-specific features. We have also considered the associations that exist among the set of possible labels. We have used hinge loss function which ensures less sensitivity towards out-liers and the accelerated proximal gradient method (APG) to efficiently solve the underlying optimization problem. Our proposed approach is competitive with other state-of-the-art approaches according to experimental findings on several benchmark data sets.(c) 2022 Elsevier B.V. All rights reserved.
Keywords:
Multi-label learning
Laplacian matrix
Nearest neighbor
Accelerated proximal gradient