LabCor: Multi-label classification using a label correction strategy

Multi-label classification is a branch of machine learning that can effectively reflect real-world problems. Among all the multi-label classification methods, stacked binary relevance (2BR) is a classic approach. Based on 2BR, a series of optimized algorithms have been derived. Although these algorithms have adapted complex optimizing processes and shown remarkable performance, their core concepts are rather similar, mainly involving restructuring the feature spaces of meta-classifiers. Existing research has rarely discussed that the use of inappropriate two-level predictions causes negative impacts on 2BR structures. In this study, we propose a 2BR-based label correction method named LabCor, which focuses on the identification and correction of unreliable two-level predictions. We first discuss the inner mechanism by which the 2BR-based algorithms obtain their two-level outputs and find a marker that can reflect the reliabilities of the samples predictions. Based on the mechanism, we then introduce a graph based output determination method that can use the training samples to generate dimensional decision patterns. The global label count distribution is also used to reflect the goal of classification problems. In the prediction phase, LabCor uses the decision patterns and label count constraints to identify and correct the misclassified labels. According to the evaluation results, the proposed method can effectively reduce the impact of troubling two-level predictions and yield superior or competitive performance versus well-established 2BR-based algorithms.