Research Of Traditonal Chinese Medicine Inquiry Based On Multi-label Learning

The syndromes of patients were judged by asking the symptoms in inquiry of Traditional Chinese Medicine (TCM). This diagnostic process requires doctors had extensive experience of Syndrome Differentiation at the time of treatment. With the rapid development of machine learning and pattern recognition recently, many problems requiring individual judgment can be solved by building complex mathematical models. The forecasting results are completed by computers.The multi-label learning technique is a new research direction in data mining (DM). Because in reality there are strong correlations existing in features or labels, and the dependencies between features and labels are complicated. Against the single-label learning method, the multi-label learning method is no longer for the binary label classification irrelevantly. It is not only considered the correlations between labels in real conditions, but also improves the classification accuracies.However, the dependencies between features and labels in the original data sets are ignored if we only utilize the multi-label learning technique, which focus on the correlations of labels. At the result, the Feature Dimension Reduction becomes the hot issue to effectively remove some irrelevant and redundant features in multi-label learning research. The new method improves the classification accuracy further.Aimed at the topic of comparing the classification algorithms in coronary heart disease (CHD) data sets and exploiting the TCM inquiring system, this dissertation proposes a multi-label learning framework based on pairwise-class feature selection. It is called multi-label learning with Relevant fEature for eAch Label (REAL). In the process, we elaborate design concept, improve several existing multi-label learning methods and make experiments with CHD, Medical, Enron and Scene datasets. Furthermore, we optimize the feature subsets by Co-evolutionary Genetic Algorithm (CGA), which are inputted into the multi-label classifiers based on REAL framework. Finally, a TCM inquiring system is exploited based on VC++language. It basically implements the function that machine distinguishes the syndromes of patients. The main work and contributions of the paper are show as follows:1. Selection of Feature Dimension Reduction. There are a few feature selection methods used on multi-label learning algorithms, such as MLSI, MLNB, MDDM, MEFS. They are used as pretreatment process of multi-label learning classification directly. The feature set for all class labels is likely to be bias to every group of possible class labels. As a result, we choose the pairwise-class feature filter to select feature subsets from the original data set. The correlations between features and each group of class labels are fully taken into account.2. Design of a multi-label learning REAL framework. The new strategy integrates the pairwise feature filter method. The dependencies between features and labels are considered based on preserving the correlations between class labels. The combination of different feature selection methods and multi-label classifiers is used in the framework. The CMIM method is fitted into the modified ML-kNN and LIFT algorithms in the paper. Furthermore, the selected feature subsets are optimized secondly by GCA in order to filter out redundancies and retain more important information. We make experiments for each step of improvement. The classification accuracy is greatly enhanced to87.8%.3. Exploit of TCM inquiring system based on VC++language. The REAL algorithm is packaged into the TCM inquiring system based on VC++6.0. The basic machine-inquiring functions are accomplished by this software system, including the essential information entry of patients, the symptoms input and retrieve of patients, the training and test model of the differential TCM based on REAL algorithms, the statistics and analysis of syndromes and the print of inquiring record, etc.