Improvement of classification performance for a multi-label classifier by a minority label synthesis

Abstract

A multi-label dataset has multiple labels in a single input data point. Label imbalance data with an extreme bias in the frequency of occurrence of each label is likely to occur in multi-label data, and it is a challenge to improve the classification accuracy of label imbalance data. Multi-Label CIM-based ART (MLCA) is a previous research on multi-label classifier. It can continuously learn new data and label information. MLCA generates nodes from input data and uses Naive Bayes for label classification. MLCA counts the labels of the data belonging to the node in learning the label information. Since MLCA does not have sufficient countermeasures against unbalanced data, the classification performance may be impaired. One of the methods for mitigating imbalance data is MLSOL. This method can generate data not only by considering label imbalances in the data as a whole, but also by considering local label imbalances. In this study, we propose a combination of MLCA and MLSOL. We aim to improve the classification performance for imbalanced data by correcting the label count based on the data generated by MLSOL when MLCA is training on label information. Numerical experiments show that the classification performance of unbalanced data was improved compared to before label correction was applied.