关键词: 关键词: 多标签, 标签相关性, 贝叶斯网络, 信息增益, 特征选择

Abstract: Abstract: ECG samples usually exhibit many diseases, and it is a typical multi-label classification problem. In multi-label classification methods, RAKEL (random k-labelsets) method decomposes the initial label set into several random subsets of size k, and establishes LP (label powerset) classifier for training. Because of the random selection of labels, the correlation between labels is not fully considered, which can easily cause few training examples for some new label subsets, which will deteriorate the prediction performance. This paper proposes a BN-RAKEL (Bayesian-network based RAKEL) algorithm. Firstly, the correlation among labels is found by Bayesian network, and then the candidate label subset is determined by using the feature selection method based on information gain. For each candidate label subset, the correlation degree is detected by using the similarity of optimal feature space. Finally, LP classifier is trained in the optimal feature space of the label subset. The experiment shows that the improved algorithm has improved prediction performance. Among the 18 diseases included in the ECG data set, the Recall and F_score values of 17 diseases have been improved compared with the RAKEL algorithm.

Key words: Keywords: multi-label, label correlation, bayesian-network, information gain, feature selection