基于贝叶斯模型的多标签分类算法

Multi-label classification algorithm based on Bayesian model

针对二元关联法(BR)未考虑标签之间相关性,容易造成分类器输出在训练集中不存在或次数较少标签的不足,提出了基于贝叶斯模型的多标签分类算法(MLBM)和马尔可夫型多标签分类算法(MMLBM)。首先,建立仿真模型分析BR算法的不足,考虑到标签的取值应由属性置信度和标签置信度共同决定,提出MLBM。其中,通过传统的分类算法计算获得属性置信度,以及通过训练集得到标签置信度。然后,考虑到MLBM在计算属性置信度时必须考虑所有已分类的标签,分类器的性能容易受无关或弱关系的标签影响,所以使用马尔可夫模型简化置信度的计算提出了MMLBM。理论分析和仿真实验表明,与BR算法相比,MMLBM的平均分类精度在emotions数据集上提高约4.8%,在yeast数据集上提高约9.8%,在flags数据集上提高约7.3%。实验结果表明,当数据集中实例的标签基数较大时,相对于BR算法,MMLBM的准确性有较大的提升。

Since the relation of labels in Binary Relevance（ BR） is ignored, it is easy to cause the multi-label classifier to output not exist or less emergent labels in training data. The Multi-Label classification algorithm based on Bayesian Model（ MLBM） and Markov Multi-Label classification algorithm based on Bayesian Model（ MMLBM） were proposed. Firstly, to analyze the shortcomings of BR algorithm, the simulation model was established; considering the value of label should be decided by the attribute confidence and label confidence, MLBM was proposed. Particularly, the attribute confidence was calculated by traditional classification and the label confidence was obtained directly from the training data. Secondly, when MLBM calculated label confidence, it had to consider all the classified labels, thus some of no-relation or weak-relation labels would affect performance of the classifier. To overcome the weakness of MLBM, MMLBM was proposed, which used Markov model to simplify the calculation of label confidence. The theoretical analyses and simulation experiment results demonstrate that, in comparison with BR algorithm, the average classification accuracy of MMLBM increased by 4. 8% on emotions dataset, 9. 8% on yeast dataset and 7. 3% on flags dataset. The experimental results show that MMLBM can effectively improve the classification accuracy when the label cardinality is larger in the training data.