Multi-label Classification Algorithm Based On Labels’Correlation

With the development of information technology, multi-label classification methods are widely applying, however, existing multi-label algorithms do not consider the information of relationship between labels, but the information is very helpful to improve the performance of classifier, so it is necessary to conduct the thorough research to the correlation between labels.MLKNN is one classic multi-label classification algorithm which has a good performance, but the performance of it can be improved because of the missing of considering the correlation between labels. This thesis improves MLKNN by modifying the posteriori probability and considering the correlation between labels. It clusters the dataset into some clusters firstly, and then applies the improved MLkNN algorithm in each cluster. Experiment shows that the improved method significantly improved the accuracy of the algorithm.KNN classification does not require explicit rules, and it has higher classification accuracy than other algorithms. Considering the advantage of KNN, unconditional label correlations promote each other small, and the testing of the correlation between labels is complicated. In this thesis, a new algorithm has been proposed for the study of correlation, the different correlation between the labels in different k neighbors. Experiments show that it can obtain a better performance than MLKNN by applying the new method for correlation, and has analyzed the influence of algorithm by K values, concluded that the choice of casual k value is unscientific.In this thesis, the study of conditional labels correlation provides a meaning idea for the study of correlation in multiple labels.