Selective classifier chains based on max-relevance and min-redundancy for multi-label classification

The classifier chains method is one of the most well-known methods for multi-label classification, which can model label correlations while keeping acceptable computational complexity. However, the drawbacks are that potential label redundancies may be overlooked and label relevances have not been specifically measured, making a rough and redundant model. In this paper, we present a new architecture of the classifier chains based on max-relevance and min-redundancy feature selection method, called mRMR-CC, and provide a dynamic learning algorithm of selection labels in the process of classifier chains. The algorithm considers concretely not only label correlations but also label redundancies, which allows us to select a compact additional attributes set for each base classifier. A series of numeric studies are performed using a broad range of multi-label data sets with a variety of evaluation metrics. Extensive experiments show that the proposed selective classifier chains model leads to promising improvement on additional attributes selection of the classifier chains method and predictive performance.