Class-imbalance Issue In Applying Multi-label Learning To The Study Of Parkinson In Traditional Chinese Medicine Diagnosis

Parkinson’s Disease (PD) is a common chronic central nervous system degeneration disease in middle aged and elderly people. Traditional Chinese Medicine (TCM) has a long history of research on Parkinson’s disease and also has different opinions about Parkinson’s disease syndrome. Through years of experience in TCM, provisions of Parkinson’s disease described by the five kinds of syndromes. Chinese medicine believes that Parkinson patients with up to two types of syndrome which contains the primary and secondary points. In order to standardize the process of traditional Chinese medicine diagnosis of Parkinson’s disease, the modern Chinese medicine has put forward the Parkinson TCM scale, which covers the clinical symptoms associated with Parkinson’s disease. However, for how to infer the syndrome from the scale of the symptoms, the Chinese medicine is still unable to reach a consensus, the diagnosis is still dominated by experience.In this paper, multi label learning is applied to the diagnosis of Parkinson in Chinese medicine, to explore the relationship between symptoms and syndromes, trying to provide assistance for the diagnosis process. The main work of this paper:1). In order to apply multi-label learning to the field of traditional Chinese medicine diagnosis of Parkinson, the inquiry of the scale treated as attributes, corresponding syndromes as label. In the process, there is serious imbalance problem in Parkinson data set.2). Aiming at the problem of lack of data representation of little class sample in multi-label imbalance issue, an ADAptive Multi-Label Synthetic Minority Over-sampling Technique(ADA-MLSMOTE) is proposed base on ideas of the distinct contribution of the sample and the abnormal data sample filter. The algorithm solves multi-label imbalance problem well from the data level.3). Aiming at the impact of label relevance to the multi-label imbalance issue, an algorithm named Random K-labelset Under Random Ensemble sample (RKUE) is proposed, which based on the ideas of dividing the whole label set into a plurality of sub sets and integrating multi under-sampling data sets. The experimental results show that compared with the existing multi-label algorithm, the algorithm can better solve the unbalanced phenomenon.