基于标签相关性的多标签分类算法及其在帕金森诊疗领域中的应用

发布时间：2019-05-24 21:46

【摘要】：中医量表中根据各症状对相应的证型作出预测本身是个典型的多标签分类技术。故本文研究思路是将中医量表中各症状作为特征属性,每个量表对应的证型作为标签,症状到证型的推断将通过多标签分类算法得到。我们应用修正前的帕金森数据,主要工作如下:1)为了解决提出了 Classifier Chains(CC)算法中标签顺序链中存在的随机性对分类准确率的影响,我们提出一个基于CC思想的多标签分类优化算法Entropy based Classifier Chains(ECC),该算法计算出一个优良标签预测链,以此讨论帕金森证型之间的相关性。ECC认为标签的信息熵值越小,不确定性越小,则标签被正确预测的概率就越大,因此首先计算出各标签的信息熵值并选择具有最小值的标签作为链首标签。然后从该标签开始构建最小权重标签树,最后遍历该树结点形成最终标签预测链并应用于CC模型。2)在ECC的基础上加上约瑟夫环机制,提出了一个新的算法Josephus based Classifier Chains(JCC)。JCC认为ECC形成的标签预测链并非全局有序,标签之间基于相关性的排序仍然存在一定的随机性,需要借助约瑟夫环机制最大程度地降低该随机性。3)在JCC的基础上加上了一个基于惩罚机制的动态报数方法,提出一种新的多标签分类算法PeNalty based Classifier Chains(PNCC)。该算法考虑到JCC对ECC产生的标签预测链中随机性的降低程度有限,通过一种惩罚机制进一步降低随机性,并产生最终的标签预测链应用到CC模型中。实验证明,以上三种算法均可以挖掘出一些帕金森数据集的有用信息,对于其他数据集亦有优良的准确率表现。
[Abstract]:It is a typical multi-label classification technique to predict the corresponding syndrome types according to each symptom in the scale of traditional Chinese medicine (TCM). Therefore, the idea of this paper is to take each symptom in the TCM scale as the characteristic attribute, and the syndrome type corresponding to each scale as the label, and the inference from the symptom to the syndrome type will be obtained by the multi-label classification algorithm. Using the modified Parkinson's data, the main work is as follows: 1) in order to solve the influence of randomness in the label sequence chain in the Classifier Chains (CC) algorithm on the classification accuracy, We propose a multi-label classification optimization algorithm based on CC, which calculates an excellent label prediction chain and discusses the correlation between Parkinson's syndrome types. The smaller the uncertainty is, the greater the probability that the label will be predicted correctly. Therefore, the information entropy value of each label is calculated and the label with the minimum value is selected as the chain head label. Then the minimum weight label tree is constructed from the tag, and finally the node of the tree is traversed to form the final label prediction chain and applied to the CC model. 2) on the basis of ECC, the Joseph ring mechanism is added. In this paper, a new algorithm, Josephus based Classifier Chains (JCC). JCC, is proposed, which holds that the label prediction chain formed by ECC is not globally ordered, and there is still a certain randomness in the ranking based on correlation between tags. It is necessary to minimize the randomness with the help of Joseph ring mechanism. 3) on the basis of JCC, a dynamic report method based on penalty mechanism is added, and a new multi-label classification algorithm PeNalty based Classifier Chains (PNCC). Is proposed. The algorithm takes into account the limited degree of reduction of randomness in the label prediction chain generated by JCC to ECC, further reduces the randomness through a punishment mechanism, and produces the final label prediction chain to be applied to the CC model. The experimental results show that the above three algorithms can mine some useful information of Parkinson's dataset, and also have excellent accuracy performance for other datasets.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R742.5;TP311.13

【参考文献】