.png){kind=icon}
.png){kind=icon}
Return
An improved two-stage binary relevance method for multilabel classification
.png){kind=icon}
PRE
AI
DOI：10.1080/02664763.2025.2567977.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multilabel classification concerns a family of unconventional classification problems, where each instance may be associated with multiple labels simultaneously. One of the traditional methods for multilabel classification is the binary relevance algorithm, in which one first converts the multilabel dataset into multiple binary-outcome datasets, one for each label, then trains a binary classifier on every transformed dataset and predicts the occurrence of the given label, and in the end aggregates all predicted labels together to compose a multilabel outcome. Two significant drawbacks of this method include its failure to account for label correlation and its poor predictive performance when the labels are sparse. To address these pitfalls, we developed an improved two-stage binary relevance method that utilizes cluster analysis to unveil underlying label structures and subsequently realizes label subset and individual label predictions in a sequential manner. In an effort to alleviate the challenge resulted from label subset imbalance, we consider several refined realizations of the proposed methodology. We showcase the proposal on ten real multilabel datasets and compare it to the benchmark binary relevance method as well as three other competing methods using five different performance metrics. The numerical studies reveal that the proposal yields significantly better or comparable results to its counterparts in almost all cases.
Keywords:
Binary relevance
clustering
label correlation
multilabel classification
sequential prediction
Journal
J
IF:
1.1
Papers:
131
Citations:
4.2K