An Improved Multi-label Classification Algorithm BRkNN

摘要

Binary Relevance algorithm is widely used in multi-label classification. However, this algorithm ignores the correlation between different labels, so when the label set is in large scale, the algorithm shows poor performance. To tackle this issue, this paper proposes an improved BRkXN multi-label classification algorithm (BRkNN-new). The algorithm not only computes the distance similarity between test instances and training instances, but also infuses the adjacent training labels' co-occurrence into final decision. In order to add the label correlation information into the BRkNN algorithm, this paper also proposes a label transformation method, which is based on the principle that labels with much more similarity have more possibility to be classified into one class finally. First the initial clustering is realized by BRkNN and the distance similarity is measured between test instances and training instances, and then computes the prior information of the label correlation from the training samples, and transfers the correlation information among the labels of the initial clustering k-nearest neighbors Finally the distance similarity and label correlation information are integrated to get the final classification result. The experiment result suggests that the BRkNN-new algorithm can make up the defect of the existing BRkNN algorithm and shows good performance in scene, emotions and yeast datasets.